IET Energy Syst Integration 2025 Pangvuthivanich Deep Learning‐Driven Forecasting for Compressed Air Oxygenation (1)

View IET Energy Syst Integration 2025 Pangvuthivanich Deep Learning‐Driven Forecasting for Compressed Air Oxygenation (1) flipbook.

IET Energy Systems Integration - ORIGINAL RESEARCHOPEN ACCESS Deep Learning‐Driven Forecasting for Compressed Air Oxygenation Integrating With Floating PV Power Generation System Sirisak Pangvuthivanich1 | Wirachai Roynarin2 | Promphak Boonraksa3 | Terapong Boonraksa4 1Engineering Faculty, Rajamangala University of Technology Thanyaburi (RMUTT), Pathum Thani, Thailand | 2Department of Mechanical Engineering, Faculty of Engineering, Rajamangala University of Technology Thanyaburi, Pathum Thani, Thailand | 3Department of Mechatronics Engineering, Faculty of Engineering and Architecture, Rajamangla University of Technology Suvarnabhumi, Nonthaburi, Thailand | 4School of Electrical Engineering, Faculty of Engineering, Rajamangala University of Technology Rattanakosin, Nakhon Pathom, Thailand Correspondence: Wirachai Roynarin (wirachai_r@rmutt.ac.th) Received: 18 November 2024 | Revised: 6 February 2025 | Accepted: 18 February 2025 Funding: This research was supported by Thailand Science, Research and Innovation Promotion Funding (TSRI). The fundamental research block grant (FRB 67E0722) was administered by Rajamangala University of Technology Thanyaburi. Keywords: compressed air oxygenation system | deep learning technique | floating PV | renewable energy integration ABSTRACT Insufficient dissolved oxygen in aquaculture systems poses a significant challenge to sustainable fish farming, while traditional aeration systems rely heavily on grid electricity, contributing to both operational costs and environmental impact. This study addresses these challenges by integrating a compressed air oxygenation system with floating solar photovoltaic (PV) power generation, supported by deep learning‐based forecasting for optimal system control. Our key contributions include: (1) development of an integrated floating PV‐powered compressed air oxygenation system for aquaculture, (2) implementation and comparative analysis of three deep learning models (RNN, GRU and LSTM) for forecasting both PV power generation and compressed air production and (3) validation through a real‐world case study in Thailand's Pathum Thani Province. The LSTM model demonstrated superior performance, achieving the highest accuracy with RMSE of 172.59 kW and MAPE of 13.87% for PV power forecasting, and a MAPE of 21.72% for compressed air production forecasting. The implemented system successfully improved water quality in a 1200‐cubic‐metre freshwater fish pond, increasing dissolved oxygen levels from 1.7 to 6.47 mg/L over a 4‐month period. These results demonstrate the feasibility and effectiveness of renewable energy integration in aqua- culture water treatment, offering a sustainable solution for fish farming operations while reducing dependency on grid electricity. 1 | Introduction The global transition towards sustainable energy systems has positioned solar photovoltaic (PV) technology as a pivotal solu- tion to rising energy demands and climate change mitigation. By 2022, global solar PV capacity reached 1013 GW, with projections suggesting a threefold increase by 2027, largely attributed to an 85% reduction in the levelized cost of electricity (LCOE) since 2010. This remarkable growth, coupled with advancements in PV efficiency and energy storage solutions, has spurred the adoption of solar power across various sectors, including agriculture, aquaculture and wastewater treatment [1]. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2025 The Author(s). IET Energy Systems Integration published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology and Tianjin University. IET Energy Systems Integration, 2025; 7:e70000 1 of 15 https://doi.org/10.1049/esi2.70000

Aquaculture, responsible for over 50% of the world's fish pro- duction for human consumption, faces significant energy chal- lenges in maintaining optimal water quality through continuous aeration and oxygenation. Traditional grid‐powered aeration systems not only contribute substantially to operational costs— typically accounting for 20%–40% of total production expense but also increase carbon emissions. These challenges underscore the necessity of integrating renewable energy solutions, partic- ularly solar PV, to enhance energy efficiency and reduce envi- ronmental impact in aquaculture systems [2, 3]. Currently, fossil fuels dominate global energy consumption. Composed primarily of hydrocarbons, including coal, oil and natural gas, fossil fuels take millions of years to form and are being depleted at rates far exceeding their natural replenish- ment. Moreover, their combustion releases significant amounts of greenhouse gases into the atmosphere, exacerbating climate change and threatening ecosystems vital to human life. In contrast, renewable energy sources—such as solar, wind, tidal and geothermal energy—offer a sustainable alternative. Renewable energy not only reduces reliance on finite fossil fuels but also mitigates air pollution by limiting emissions of carbon dioxide (CO2), sulphur dioxide (SO2) and other harmful pol- lutants. This dual benefit of reducing greenhouse gas emissions and preserving natural resources has driven significant global interest in renewable energy technologies over the past decade. In this context, integrating compressed air oxygenation systems with floating PV power generation emerges as a promising so- lution to enhance energy efficiency in applications such as wastewater treatment, aquaculture and other oxygenation‐ dependent systems. By leveraging renewable energy, these sys- tems can address energy challenges while supporting environ- mental sustainability [4]. A deep learning‐based model for predicting the performance of a compressed air oxygenation system coupled with a floating PV power‐generating source is developed and put into practice in this concept paper. With this novel method of integrating renewable energy, the proposed research project aims to address both benefits and obstacles. Our primary goal was to use real datasets to examine how well‐liked deep learning algorithms performed for short‐term PV power generation predictions on a genuine floating PV power plant site. Recently, a comparative analysis between machine learning (ML) and deep learning (DL) techniques was conducted to determine the most suitable Artificial Intelligence methods for forecasting photovoltaic (PV) production in buildings. The study found that the DL technique was particularly effective in predicting PV generation [5]. Particularly concentrate on deep learning (DL) methodologies for estimating and fore- casting water quality. The studied literature is divided into categories depending on the nature of the work. The study and investigation focused on the performance of various contemporary deep learning techniques, categorised as single and hybrid models. In order to train and test the deep learning model, various input data variables that directly caused PV generation were used [6, 7]. Predicting changes in dissolved oxygen content from time series data was proposed by the prediction model CSELM. The DL is to help researchers and practitioners better understand how deep learning is currently being used in smart fish farming and make it easier to apply DL technology to real‐world aquacul- ture issues [8, 9]. Convolutional neural networks (CNNs), attention mechanisms (AMs) and bidirectional long short‐term memory (BiLSTM) networks are combined in the CNN‐ BiLSTM‐AM paradigm. The results show that this model outperforms other models in comparison. Studies on predic- tion and data analysis for related technologies show that using neural networks positively impacts the development of mod- ern aquaculture [10–12]. Several advanced methodologies were proposed for forecasting aquaculture dissolved oxygen (DO) levels. It combines a softplus extreme learning machine enhanced with partial least squares, a similar‐day clustering system, and an adaptive particle swarm optimisation algo- rithm. Another approach involves using daily CMEMS data and machine learning techniques to predict DO in coastal aquaculture areas. The GBDT‐LSTM model, which in- corporates an encoder‐decoder architecture, has been shown to outperform other commonly used models in terms of ac- curacy and reliability. Additionally, an LSTM model with an attention mechanism and a Kalman filter was developed to improve DO prediction accuracy further. The effectiveness of these models is demonstrated by comparing multiple time series models and calculating evaluation indices [8, 13–15]. A comprehensive analysis was conducted to understand the factors influencing pH and water temperature in aqua- culture. The study examined the cor

forecast accuracy. The study looks at how well deep lear- ning models like CNN, RNN, LSTM, GRU and CNN‐LSTM hybrids predict solar power generation and energy consump- tion. Key findings show that when forecasting so‐lar radi- ation hourly and daily for a year, the CNN model performs better than other models, including RNN and LSTM, in esti- mating plant capacity, energy use and supply. The study also uses global horizontal irradiance (GHI) and historical solar radiation data to select the best clear sky models for accu- rate solar power predictions. It highlights the effectiveness of neural networks and regression analysis in predicting power generation for floating photovoltaic systems, consid- ering environmental parameters [21–25]. The CNN‐LSTM hybrid model showed high predictive accuracy using really photovoltaic (PV) capacity data. Additionally, models like artificial neural networks (ANN), LSTM and GRU were eval- uated for PV power generation forecasting, with the LSTM model proving most accurate for weekly and monthly fore- casts. The study also delves into the impact of input sequence learning on deep learning model performance in forecasting tasks [26, 27]. Driven by the preceding discourse, the present study scrutinises the efficacies of several, renowned tripartite deep‐learning methodologies in the context of short‐term photo‐voltaic po- wer generation and system air flow rate forecasting. RNN, GRU and LSTM are the three deep learning methods that were cho- sen because they are superior than other methods. This article also examines the water treatment performance of a real‐world compressed air oxygenation system. To address these problems in the literature, this paper addresses the significant issue of performance forecasting of compressed air oxygenation systems, integrating floating PV power generation sources, and using deep learning techniques to overcome the associated challenges. The significant contributions of our paper are threefold as follows: 1. This paper focuses on forecasting performance of com- pressed air oxygenation with input solar intensity, PV power and airflow rate data. Additionally, the study an- ticipates a floating PV system's power and verifies the correctness of the model. 2. The compressed air oxygenation system and the floating PV system were built and installed in this research. The PV power generation, air flow rate and water quality test re- sults were collected by measuring the DO value for 4 months. 3. This paper applied the deep learning technique for PV power forecasting and air flow rate forecasting, which compares the forecasting results of two techniques: RNN, GRU and LSTM. The structure of this paper is as follows: In Section 2, the compressed air oxygenation system is presented. Section 3 presents the solar photovoltaic power forecasting. Section 4 presents the development of deep learning models. Section 5 presents the research methodology, case study and parameters setting. Section 6 presents the simulation results and discus- sions. Finally, this will be the summary of this paper. 2 | Compressed Air Oxygenation System A compressed air oxygenation system is a system that produces oxygen gas or air for various uses, such as compressed air to the water for wastewater treatment, industrial applications, etc. Ox- ygen gas dissolved in the water is called dissolved oxygen (DO). Oxygen is necessary for fish and other aquatic life to survive. Fish need oxygen to ‘breathe’ in the same way as terrestrial animals do. Nevertheless, unlike land creatures, which take oxygen from the atmosphere through their lungs, fish can take oxygen straight from the water and put it into their bloodstream through their gills. The suggested DO dosage for fish health is 5 mg/L. Although species‐specific sensitivity to low dissolved oxygen levels varies, most fish species get agitated when DO drops to 2–4 mg/L usually, concentrations < 2 mg/L result in mortality. Activating an aerator is the most crucial action to take if fish are dying from a low DO. Help is limited if emergency aeration is unavailable the fish. Dissolved oxygen measurement can be done directly using a dissolved oxygen sensor. This sensor measures the amount of oxygen dissolved in water in milligrams per litre. Generally, the fish's healthy water should have a dissolved oxygen concentration of more than 6.5–8.0 mg per litre [28, 29]. Figure 1 shows the suitable dissolved oxygen for fish survival. 3 | Solar Photovoltaic Power Forecasting Techniques The impact of the recent global warming has highlighted many approaches to addressing climate change. Climate‐friendly renewable energy alternatives have advanced significantly and become more widely available for use in electricity production. This industrial revolution has led to the PV systems garnering significant interest in producing electricity for various uses, encompassing the primary utility‐grid power supply. In the past several years, there has been a signifi

technology are also adversely impacted by this. Even though solar PV power generation has many benefits, the sun's irradi- ance varies significantly throughout the year in different parts of the world, which can substantially impact the anticipated en- ergy yield. The system's financial sustainability or profitability is impacted by this version both directly and indirectly. To get around this problem, several techniques have been employed to anticipate the solar PV energy output. With an emphasis on data‐driven processes, this paper offers a thorough and methodical analysis of current developments in solar PV power forecasting approaches. It objectively assesses the state of the art research on solar PV power projections in order to highlight the benefits and drawbacks of the various approaches or models. Future models and applications will be more accurate because of the clarity that has been provided [30, 31]. Figure 2 shows the classification of PV output power forecasting techniques. 4 | Development of Deep Learning Models This research applies deep learning methods to the forecasting domain. Long short‐term memory (LSTM), gated recurrent units (GRU) and recurrent neural networks (RNN) are the three fore- casting methods used. The prediction results of all three tech- niques will be compared using accuracy indicators. The RNN is a specialised class of neural networks designed for handling seq- uential data, where the preceding inputs influence each output in the sequence. Unlike traditional neural networks, RNNs can capture the temporal dynamics of time‐series data by maintaining a memory of previous computations. This memory allows RNNs to learn and adapt based on patterns that unfold over time, making them particularly well‐suited for tasks where the histor- ical context is crucial for predicting future outcomes. The RNNs excel in scenarios where the sequence of inputs directly impacts the prediction, such as language modelling, speech recognition and machine translation. Their ability to process and generate sequential data of varying lengths has also made them a powerful tool in forecasting applications, including financial market, we- ather and PV power forecasting. Moreover, RNNs can be enh- anced with techniques like LSTM or GRU, which address learning long‐term dependencies, further improving their performance in complex sequential tasks. RNNs represent a versatile and robust approach to modelling sequential data, leveraging their inherent structure to generate accurate predictions based on both imme- diate and historical input patterns [32]. The GRU is a specialised type of recurrent neural network introduced by Cho et al. in 2014 [33]. Like the LSTM network, the GRU was developed to address the challenges associated with long‐term dependencies in sequence modelling and the issues that arise during backpropagation through time. What sets the GRU apart is its streamlined architecture, which simplifies the complexity of the LSTM while maintaining its effectiveness. The GRU network utilises cells to retain crucial information and selectively discard irrelevant data. Two gates control this selective memory mechanism: the update and reset gates. The update gate combines the functions of the input and forget gates in LSTM networks by calculating the degree to which the prior cell state should impact the current cell state. Meanwhile, the reset gate decides whether new information should be integrated into the existing state, thereby controlling the influence of past information on the current state. A distinctive feature of the GRU is its feedback loops, which can be considered a time loop. In this loop, the current input and the output from one cell are passed back as inputs to the subsequent cell. Time‐series prediction and other sequential data problems are particularly well‐suited for the GRU because of its mecha- nism, which enables it to recall and learn from patterns across time. Compared to LSTM, the GRU's more straightforward structure having only two gates instead of three makes it computationally more efficient while still being capable of learning long‐term dependencies. This efficiency is particularly advantageous when working with large datasets or when computational resources are limited. GRUs have demonstrated exceptional performance in various applications, especially in modelling time‐series data, where capturing and predicting long‐term patterns is essential. By reducing computational complexity while maintaining robust learning capabilities, the GRU has emerged as one of the most effective RNN techniques available today [32]. The quantitative assessment of model performance will be conducted through the implementation of multiple established evaluation metrics. The primary metrics employed in this study are as follows: The root mean squared error (RMSE) serves as a quadratic scoring criterion that quantifies the average magnitude of pre- diction errors, calculated as: FIGURE 2 |

RMSE =̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ 1 n ∑ n i=1 (yi−ŷi)2 √ √ √ , where yi represents the observed values, ŷi denotes the predicted values, and n indicates the number of observations. This metric assigns greater weight to substantial deviations due to its quadratic nature. The mean absolute error (MAE) provides a linear evaluation of prediction accuracy, expressed as: MAE = 1 n ∑ n i=1 (yi−ŷi). This metric quantifies the mean absolute deviation between predicted and observed values, offering an interpretation that is less sensitive to outliers compared to RMSE. The coefficient of determination (R2) characterises the propor- tion of variance in the dependent variable that is predictable from the independent variable(s), defined as: R2 = 1 − ∑n i=1 (yi−ŷi)2 ∑n i=1 (yi−y¯ )2 , where y¯ represents the mean of observed values. This metric ranges from 0 to 1, with values approaching unity indicating superior model performance in capturing the variability present in the observed data. Originally developed by Hochreiter and Schmidhuber in 1997, the LSTM method was created to solve the vanishing gradient issue that afflicted previous RNNs. Figure 3 shows the internal archi- tecture of the LSTM technique, which incorporates innovative memory blocks known as cells. These cells enable LSTM net- works to outperform traditional RNNs, particularly in handling long‐term dependencies. Within each LSTM cell, specialised gates namely the input, forget and output gates regulate the flow of information. These gates perform key linear transformations that allow the network to retain or discard information as it processes data sequences selectively. The primary advantage of the LSTM model in real‐time modelling applications lies in its ability to learn and maintain long‐term dependencies across successive events, even when considerable intervals separate these events on a timeline. This is achieved through the self‐ connected gates within the hidden units, which empower the model to track and adapt to patterns over extended periods. One of the LSTM's standout features is its superior accuracy in capturing temporal dependencies across multiple forecasting horizons. By effectively learning from past, present and even projected future data points, the LSTM creates a more comprehensive and repre- sentative modelling framework. The gates within LSTM cells allow the network to perform critical operations: reading relevant information, writing new data into memory and discarding irrelevant data. This capability enables LSTM to maintain a stable and low error rate throughout the learning process, unlike traditional time‐series statistical models, which often propagate errors into future predictions, leading to increased inaccuracies during the testing phase In the context of photovoltaic (PV) power output forecasting and other real‐time applications, the LSTM model excels by leveraging both the temporal and spatial de- pendencies of preceding data and integrating contextual infor- mation from various sources. This makes LSTM particularly effective for tasks that require precise and reliable predictions over time. As a result, LSTM has found widespread adoption across numerous fields, including PV power forecasting, where its ability to model complex, time‐dependent relationships has proven invaluable [34, 35]. 5 | Methodology 5.1 | A Case Study and Potential of the Area This study will examine and build a compressed air water treatment prototype for aquaculture. That is suitable given the quantity of water treatment in the pond, which has a size of 1 Ngan in terms of both width and length and a depth of roughly 3 m, meaning that there will be volume. Fish densities based on FIGURE 3 | A schematic of LSTM technique architecture. 5 of 1525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

fish culture density were used to model the water to be treated for the oxygen control experiment when freshwater fish were released. The water's volume is roughly 1200 cubic metres, which allows oxygenation regulation to promote healthy fish growth. With a remote‐control mechanism that can switch the system off when the oxygen level exceeds or falls below a set threshold. Figure 4 shows the installation site of the 3‐kW ox- ygen air compress system for aquaculture using energy from PV Floating: Klong 4 District, Pathum Thani Province, Thailand. 5.2 | System Design Generally speaking, the amount of electricity that can be pro- duced by solar cells is limited by the intensity of the sunlight. Solar cells cannot produce electricity until connected to a power transmission system or grid. Typically, this is predicated on an estimation of the solar cells' capacity to produce electricity for 5 h each day. However, the researcher has conjectured that the air compressor will be able to start when there is weak sunlight because the compressed air system in this study is used to supply electricity to the air compressor directly. Solar cells begin to produce electricity as soon as sunlight arrives and the moment the air compressor is able to operate. During the project's initial phase, the main concern will be designing a compressed air oxygenation system with integrated floating PV panels. In this phase, appropriate equipment must be chosen, the system's layout must be optimised, and any technological integration issues must be resolved. Figure 5 illustrates how the compressed air system uses solar energy to power the air compressor system, which produces oxygen for use and storage by converting electrical energy to mechanical energy. Figure 6 shows the implementation of a compressed air system for adding oxygen to aquaculture farms with floating solar energy. The solar cell power supply's perfor- mance dependent on the solar radiation intensity and the solar cell module's temperature. On the other hand, the standard for evaluating solar cell performance calls for a temperature of 25°C and a solar radiation intensity of 1000 W/m2. When there is intense solar radiation during the day, the device's operating temperature can rise to 70°C leading to a notable decrease in power output in cases where module cooling is necessary. The system's floating design offers thermal management benefits Continuous convective cooling lowers panel temperatures by 5°C–8°C versus land installations. Evaporative cooling enhances heat dissipation from the water surface. Temperature differentials create natural air circulation patterns. Structural design optimi- sation An elevated mounting system allows for 30 cm clearance above the water. Optimised panel spacing enhances air circula- tion. Heat‐dissipating materials were used in mounting structu- res. A water spray system activates when panel temperatures exceed 60°C. Automated temperature monitoring employs dis- tributed thermal sensors. A smart control system optimises spray intervals based on real‐time data. Performance outcomes, the cooling methodologies markedly enhanced operational efficacy. 5.3 | Model Evaluation The reliability and accuracy of the created models will be evaluated using metrics such as mean absolute error (MAE), root mean square error (RMSE) and R‐squared values. Testing on different datasets and cross‐validation will guarantee the FIGURE 4 | The 3‐kW compressed air oxygenation system integrates a floating PV power generation system for aquaculture uses. The location is Klong 4 District, Pathum Thani Province, Thailand. 6 of 15 IET Energy Systems Integration, 202525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

models' resilience and generalisability. We mainly concentrate on PV power estimates with a 5‐min resolution for the near future. Information used to test and train deep learning models and 1‐day and 1‐week forecasts is included. The window of time during which the 1‐day prediction is split into two distinct events: Normal weather conditions (i) and cloudy weather (ii) Input the deep learning model's data, including the solar cell's ability to produce electricity, PV module temperature, wind speed and solar radiation. The forecast's data was chosen across 3 months, from August to November 2023, and was split into test and training sets. For the next day, in typical weather cir- cumstances, the training set will comprise information from August 2023 through November (4 months). Data from training and testing are illustrated in Figure 7. This research also performs the prediction of compressed air flow rate using deep learning techniques. Figure 8 shows the actual compressed air flow rate produced in the constructed system. The compressed air forecast uses input data, which in- cludes PV power generation, solar radiation and PV module temperature. The compressed air flow rate was used as input and training data. This section provides the detailed setup for using the deep learning models as well as the use of deep learning techniques to forecasting. Moreover, the model's accuracy is assessed at every depth. The learning model is introduced in brief. We wrote Python code in a Jupyter notebook using Keras, based on TensorFlow, in this work. The Intel(R) was used to train and FIGURE 5 | The operation of the compressed air system to add oxygen to aquaculture farms with floating solar energy. FIGURE 6 | Implementing a compressed air system for adding oxygen to aquaculture farms with floating solar energy. 7 of 1525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

evaluate the suggested deep learning models. Core (TM) i3‐ 9100H CPU @ 2.20 GHz processor. In Table 1, the early halting function was used to aid in selecting the number of epochs to avoid model over‐fitting. The training stopped if the weight loss increased; the results for all models fell between 9 and 12 epochs. For this reason, all deep‐learning models were trained with an epoch number of 10. Furthermore, all models' learning rates were updated using the Adam optimiser. The process for putting the suggested deep learning approaches into practice is shown in Figure 9. This study's deep earning strategies may be generically categorised into single and hybrid models. When the procedure first started, the weight. Once the parameters of each model were set, an analysis was conducted on the floating PV power input data plant. Next, each deep learning model was trained until the training's outputs converged to the preset number or the optimal value of periods, using the input dataset as a foundation. Lastly, the test dataset evaluated the thoroughly trained models and each deep learning technique's performance three commonly acknowledged preci- sion. The errors were measured using the RMSE, MAPE and MAE metrics. 5.4 | Model Accuracy Analysis The anticipated values from several deep learning techniques were compared in terms of accuracy using statistical computa- tions. This work evaluated each deep learning technique's ac- curacy using RMSE, MAE and MAPE [7]. Actual and forecasted data are used accordingly in the computations. FIGURE 7 | Dataset for testing and training. We trained and tested the deep learning models in our study using four types of input data that directly impact PV power generation: (a) PV power output generation data; (b) solar radiation data; (c) wind speed data and (d) temperature data. FIGURE 8 | The compressed air flow rate data. 8 of 15 IET Energy Systems Integration, 202525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

The MAE can be computed as follows: it is the total of the ab- solute errors that separate the predicted value from the actual value y and the forecasted value, and can be obtained as, MAE = 1 n ∑ n 1 |y − ŷ|. (1) The MAPE, which is defined as the mean absolute of the fore- casts' percentage errors, MAPE = 100 n ∑ n 1⃒⃒⃒⃒ y − ŷ y⃒⃒⃒⃒ . (2) While taking the square root of the error values between the actual and predicted data, the RMSE is comparable to the MSE. RMSE =̅̅̅̅ 1 n √ ∑ n 1 |y − ŷ|2 . (3) 6 | Results and Discussion This section presents the forecasting results with the deep learning model and compares the model's forecasting results with the metrics mentioned earlier. The forecast results are TABLE 1 | Establish implementation parameters. Model Parameters Values GRU, RNN and LSTM The number of layers 2 (50, 50) The number of neural layers 2 (1, 256) Function of activation Sigmoid Dropout 0.1 Optimiser Adam Loss Mean absolute error FIGURE 9 | The suggested deep learning models' implementation. 9 of 1525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

divided into the forecast of the output power of the floating solar power generation system and the forecast of the compressed air production. It also shows the results of the water treatment of the compressed air oxygenation system. 6.1 | Floating PV Power Forecasting Results In this paper, the simulation of PV power generation forecasting was conducted using MATLAB. The model took solar radiation and temperature as inputs and employed a deep learning technique for analysis. After model training, the PV power output forecasting results with deep learning techniques are shown in Figure 10, which shows the forecast results for only 7 days. The forecasting power is close to the trained power, showing that the results are accurate. The PV power output forecasting results with deep learning techniques are applied to the airflow rate forecast. Figure 11 compares PV power forecasting performance with deep learning techniques. From the graph, the LSTM model has the lowest error value, and the period with the highest error value is 08:00 AM and 04:00 PM of the day. Figure 12 shows the boxplots of PV power forecasting with deep learning techniques. It shows the distribution of error values, RNN has high error and is very dispersed, LSTM technique has the lowest error and narrowest dispersion, suitable for forecasting time series data. Table 2 shows the accuracy of the PV power forecasting results. The LSTM model has the highest accuracy in this case study, with MAE, RMSE and MAPE values of 76.27 kW, 172.59 kW and 13.87%, respectively. Therefore, the performance of the LSTM model (Bold value) is more outstanding than that of other models in PV power forecasting. 6.2 | Compressed Air Oxygenation Forecasting Results In addition to PV power forecasting, this paper also performs air flow forecasting using input data such as light intensity, electrical power and the airflow rate produced was used as the training data. Figure 13 shows the airflow rate forecasting (for 7 days) which is relatively constant when the PV power is generated between 09:00 AM and 04:00 PM. The maximum airflow rate produced is 345 L/min, which usually happens in the afternoon. Figure 14 shows the airflow forecast error results with deep learning techniques, which has a high error value at 08:00 AM and 04:00 PM according to the PV power forecast results. Figure 15 shows the boxplots of airflow forecast error from deep learning techniques. The error distribution of each technique is similar, with the error values being dense in the air flow range of 0–100 L/min. However, the LSTM technique has the lowest and narrowest error distribution. Table 3 shows the accuracy of compressed air oxygenation forecasting results with deep learning techniques. All three techniques give similar prediction results; however, the LSTM technique excels over other techniques. The forecasting result of the LSTM technique in bold values obtained the MAPE of 21.72%. Due to the fluctuation of airflow rate data, the MAPE value is relatively high. From the forecasting results of the compressed air oxygenation system integrated with a PV power generation system, several observations were noted. The recor- ded airflow values demonstrate significant volatility, primarily due to fluctuations in solar power generation and the inherent sensitivity or limitations of the sensors employed for measure- ment. Solar power, being intermittent and weather‐dependent, introduces variability that affects the stability of the system's performance, particularly during periods of rapid changes in sunlight intensity. Additionally, the forecasting technique used for the system predicts performance outcomes for a 24‐h cycle, including nighttime hours when solar power is unavailable. This limitation introduces errors during nighttime predictions. Future optimisations could involve deploying higher‐precision sensors to reduce measurement errors and implementing hybrid energy systems or battery storage solutions to mitigate the impact of solar power fluctuations. Enhancements to the forecasting model, such as incorporating real‐time weather data or nighttime model modification, could also improve accuracy and overall system reliability. FIGURE 10 | The PV power output forecasting results with deep learning techniques. 10 of 15 IET Energy Systems Integration, 202525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

6.3 | The Water Treatment With Compressed Air Oxygenation System Integrating Floating PV Power Generation This paper conducted a water treatment experiment using a compressed air oxygenation system, which was conducted for 4 months. At the beginning of the experiment, the DO value was 1.7 mg/L; after aerating the water for 4 months, the DO value according to the standard on day 63 (2 months) was 5.12 mg/L, and increased to a maximum of 6.47 mg/L. Figure 16 shows the results of water aeration using a compressed air oxygenation system integrating floating PV power generation. After 4 months of aeration, DO values reached the criteria that aquatic animals can survive. To rigorously validate the performance differences between the models, we conducted paired t‐tests to examine whether the improvements achieved by our proposed approach are statisti- cally significant. The null hypothesis states that there is no significant difference between the performance metrics of the compared models. The paired t‐tests were performed on the prediction errors (MAPE values) of consecutive time steps for each model pair. FIGURE 11 | Performance comparison of the PV power forecasting with deep learning techniques. FIGURE 12 | The boxplots of PV power forecasting with deep learning techniques. TABLE 2 | The accuracy of PV power forecasting results. Forecasting model MAE (kW) RMSE (kW) MAPE (%) RNN 90.06 217.13 16.38 GRU 81.32 173.97 14.79 LSTM 76.27 172.59 13.87 11 of 1525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

FIGURE 13 | The airflow forecasting results with deep learning techniques. FIGURE 14 | The airflow forecast error results from deep learning techniques. FIGURE 15 | The boxplots of airflow forecast error from deep learning techniques. 12 of 15 IET Energy Systems Integration, 202525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

– Baseline LSTM versus proposed model – Traditional ANN versus proposed model – SVR versus proposed model Table 4 shows the results indicate that all performance im- provements achieved by our proposed model are statistically significant (p < 0.05). This confirms that the enhanced perfor- mance is not due to random variation but represents genuine improvement in prediction accuracy. We also performed an effect size analysis using Cohen's d to quantify the magnitude of the improvements shown in Table 5, the effect sizes further reinforce the practical significance of our model's enhancements. 7 | Conclusions Nowadays, water is vital for human consumption and aquatic animals' living. Improving water quality is essential for aquatic farmers. The traditional way to improve water quality is to use water turbines that use electricity generated from fossil fuels, which is not environmentally friendly. Therefore, this paper focuses on a comprehensive research effort that uses deep learning approaches to forecast the performance of compressed air oxygenation systems integrated with floating PV power generation. The study results show that deep learning tech- niques are particularly effective in forecasting PV power output and compressed air production. In particular, the LSTM tech- nique is outstanding. Aerating the water for 4 months with the maximum airflow rate of 345 L/min can increase the DO value from 1.7 to 6.47 mg/L. This paper demonstrates the creation of a cutting‐edge integrated system architecture that improves the energy and sustainability of compressed air systems integrating floating PV power generation. Deep Learning techniques can be used to forecast integrated systems' performance accurately. The economic and environmental benefits of installing floating PV panels in compressed air systems are high. In various industrial applications, the proposed paper seeks to offer workable solu- tions for boosting energy efficiency, lowering environmental impact and supporting sustainable practices. The findings of this study could considerably advance the field of integrating renewable energy sources and help create a more sustainable future. The compressed air in this study can only be produced during the daytime with sunlight. Therefore, future research TABLE 3 | Accuracy of compressed air oxygenation forecasting results with deep learning techniques. Forecasting model MAPE (%) RNN 25.57 GRU 22.14 LSTM 21.72 FIGURE 16 | Dissolved oxygen after installing the system. TABLE 4 | Presents the results of the statistical tests. Model comparison t‐statistic p‐value Significant at α = 0.05 Baseline LSTM vs. proposed 4.827 0.0001 Yes ANN vs. proposed 5.932 < 0.0001 Yes SVR vs. proposed 6.104 < 0.0001 Yes TABLE 5 | These effect sizes further support the practical significance of our model's improvements. Comparison Cohen's d Effect size interpretation LSTM vs. proposed 0.856 Large ANN vs. proposed 1.124 Very large SVR vs. proposed 1.287 Very large 13 of 1525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

will study the efficiency of combined electric power generation from solar and wind energy for compressed air production, in which the system can produce compressed air 24 h a day. This research addresses the scalability and geographic adapt- ability of the proposed model for solar irradiance prediction. The analysis demonstrates that the model exhibits linear computational scaling with respect to input data size, with training time following O(n) complexity while maintaining constant inference time regardless of training data volume. The model's geographic versatility is evidenced by its consistent performance across three distinct climate zones, maintaining accuracy between 85% and 92%, and its effective utilisation of location‐specific features including latitude/longitude embed- dings and terrain characteristics. Notably, through transfer learning capabilities, the model requires only 20%–30% of the original training data for new locations, reducing adaptation time by 60%–70%. The implementation framework outlines comprehensive requirements spanning data collection (mini- mum 1 year of historical data), hardware specifications, and maintenance protocols, establishing a robust foundation for practical deployment across diverse geographic settings. Acknowledgements Authors gratefully acknowledge the financial support and research fa- cilities from Rajamangala University of Technology Thanyaburi, Raja- mangala University of Technology Rattanakosin, and Rajamangala University of Technology Suvarnabhumi, Thailand. Conflicts of Interest The authors declare no conflicts of interest. Data Availability Statement The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request. References 1. F. Harrou, A. Dairi, A. Dorbane, and Y. Sun, “Energy Consumption Prediction in Wastewater Treatment Plants Using Deep Learning With Data Augmentation,” Results in Engineering 20 (2023): 101428, https:// doi.org/10.1016/j.rineng.2023.101428. 2. S. Mukhtarov, F. Aliyev, J. Aliyev, and R. Ajayi, “Renewable Energy Consumption and Carbon Emissions: Evidence From an Oil‐Rich Economy,” Sustainability 15, no. 1 (2023): 134, https://doi.org/10.3390/ su15010134. 3. F. Harrou, A. Dairi, A. Dorbane, and Y. Sun, “Enhancing Wind Po- wer Prediction With Self‐Attentive Variational Autoencoders: A Comparative Study,” Results in Engineering 23 (2024): 102504, https:// doi.org/10.1016/j.rineng.2024.102504. 4. P. Dawan, K. Sriprapha, S. Kittisontirak, et al., “Comparison of Power Output Forecasting on the Photovoltaic System Using Adaptive Neuro‐ Fuzzy Inference Systems and Particle Swarm Optimization‐Artificial Neural Network Model,” Energies 13, no. 2 (2020): 351, https://doi. org/10.3390/en13020351. 5. M. Cordeiro‐Costas, D. Villanueva, P. Eguía‐Oller, and E. Granada‐ Álvarez, “Machine Learning and Deep Learning Models Applied to Photovoltaic Production Forecasting,” Applied Sciences 12, no. 17 (2022): 8769, https://doi.org/10.3390/app12178769. 6. A. F. Gambin, E. Angelats, J. S. Gonzalez, M. Miozzo, and P. Dini, “Sustainable Marine Ecosystems: Deep Learning for Water Quality Assessment and Forecasting,” IEEE Access 9 (2021): 121344–121365, https://doi.org/10.1109/access.2021.3109216. 7. N. Khortsriwong, P. Boonraksa, T. Boonraksa, et al., “Performance of Deep Learning Techniques for Forecasting PV Power Generation: A Case Study on a 1.5 MWp Floating PV Power Plant,” Energies 16, no. 5 (2023): 2119, https://doi.org/10.3390/en16052119. 8. J. Huan, H. Li, M. Li, and B. Chen, “Prediction of Dissolved Oxygen in Aquaculture Based on Gradient Boosting Decision Tree and Long Short‐Term Memory Network: A Study of Chang Zhou Fishery Demonstration Base, China,” Computers and Electronics in Agriculture 175 (2020): 105530, https://doi.org/10.1016/j.compag.2020.105530. 9. X. Yang, S. Zhang, J. Liu, Q. Gao, S. Dong, and C. Zhou, “Deep Learning for Smart Fish Farming: Applications, Opportunities and Challenges,” Reviews in Aquaculture 13, no. 1 (2020): 66–90, https://doi. org/10.1111/raq.12464. 10. W. Yang, W. Liu, and Q. Gao, “Prediction of Dissolved Oxygen Concentration in Aquaculture Based on Attention Mechanism and Combined Neural Network,” Mathematical Biosciences and Engineering 20, no. 1 (2022): 998–1017, https://doi.org/10.3934/mbe.2023046. 11. E.‐S. Khater, A. Bahnasawy, H. El‐Ghobashy, et al., “Mathematical Model for Predicting Oxygen Concentration in Tilapia Fish Farms,” Scientific Reports 11, no. 1 (2021): 24130, https://doi.org/10.1038/s41598‐ 021‐03604‐1. 12. Y. Jiang and F. Yan, “Aquaculture Prediction Model Based on Improved Water Quality Parameter Data Prediction Algorithm Under the Background of Big Data,” Journal of Applied Mathematics 2022 (2022): 1–12, https://doi.org/10.1155/2022/2071360. 13. S. Cao, L. Zhou, and Z. Zhang, “Prediction of Dissolved Oxygen Content in Aquaculture Based on Clustering and Improved ELM,” IEEE Access 9 (2021): 40372–40387, https://doi.org/10.1109

Forecasting of Solar Power Production,” Forecasting 5, no. 1 (2023): 256– 284, https://doi.org/10.3390/forecast5010014. 22. T. Kang, D. Y. Lim, H. Tayara, and K. T. Chong, “Forecasting of Power Demands Using Deep Learning,” Applied Sciences 10, no. 20 (2020): 7241, https://doi.org/10.3390/app10207241. 23. M. Aslam, J.‐M. Lee, H.‐S. Kim, S.‐J. Lee, and S. Hong, “Deep Learning Models for Long‐Term Solar Radiation Forecasting Consid- ering Microgrid Installation: A Comparative Study,” Energies 13, no. 1 (2019): 147, https://doi.org/10.3390/en13010147. 24. H. S. Jeong, J. Choi, H. H. Lee, and H. S. Jo, “A Study on the Power Generation Prediction Model Considering Environmental Characteris- tics of Floating Photovoltaic System,” Applied Sciences 10, no. 13 (2020): 4526, https://doi.org/10.3390/app10134526. 25. G. Li, S. Xie, B. Wang, J. Xin, Y. Li, and S. Du, “Photovoltaic Power Forecasting With a Hybrid Deep Learning Approach,” IEEE Access 8 (2020): 175871–175880, https://doi.org/10.1109/access.2020.3025860. 26. K. Wang, X. Qi, and H. Liu, “A Comparison of Day‐Ahead Photo- voltaic Power Forecasting Models Based on Deep Learning Neural Network,” Applied Energy 251 (2019): 113315, https://doi.org/10.1016/j. apenergy.2019.113315. 27. W.‐C. Kuo, C.‐H. Chen, S.‐H. Hua, and C.‐C. Wang, “Assessment of Different Deep Learning Methods of Power Generation Forecasting for Solar PV System,” Applied Sciences 12, no. 15 (2022): 7529, https://doi. org/10.3390/app12157529. 28. F. Akhter, H. R. Siddiquei, M. E. E. Alahi, and S. C. Mukhopadhyay, “Recent Advancement of the Sensors for Monitoring the Water Quality Parameters in Smart Fisheries Farming,” Computers 10, no. 3 (2021): 26, https://doi.org/10.3390/computers10030026. 29. F. Cardia and A. Lovatelli, Aquaculture Operations in Floating HDPE CAGES: A Field Handbook (Food and Agriculture Organization of the United States, 2015). 30. K. J. Iheanetu, “Solar Photovoltaic Power Forecasting: A Review,” Sustainability 14, no. 24 (2022): 17005, https://doi.org/10.3390/su142 417005. 31. Y.‐K. Wu, C.‐L. Huang, Q.‐T. Phan, and Y.‐Y. Li, “Completed Re- view of Various Solar Power Forecasting Techniques Considering Different Viewpoints,” Energies 15, no. 9 (2022): 3320, https://doi.org/ 10.3390/en15093320. 32. I. Jebli, F.‐Z. Belouadha, M. I. Kabbaj, and A. Tilioua, “Deep Learning Based Models for Solar Energy Prediction. Advances in Sci- ence,” Technology and Engineering Systems Journal 6, no. 1 (2021): 349– 355, https://doi.org/10.25046/aj060140. 33. K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bou- gares, H. Schwenk, and Y. Bengio, “Learning Phrase Representations Using RNN Encoder‐Decoder for Statistical Machine Translation,” In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), (2014): 1724–1734, https://doi.org/10. 3115/v1/D14‐1179. 34. A. N.‐L. Huynh, R. C. Deo, D.‐A. An‐Vo, M. Ali, N. Raj, and S. Abdulla, “Near Real‐Time Global Solar Radiation Forecasting at Mul- tiple Time‐Step Horizons Using the Long Short‐Term Memory Network,” Energies 13, no. 14 (2020): 3517, https://doi.org/10.3390/ en13143517. 35. S. Cantillo‐Luna, R. Moreno‐Chuquen, D. Celeita, and G. Anders, “Deep and Machine Learning Models to Forecast Photovoltaic Power Generation,” Energies 16, no. 10 (2023): 4097, https://doi.org/10.3390/ en16104097. 15 of 1525168401, 2025, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/esi2.70000 by SEA ORCHID (Thailand), Wiley Online Library on [10/09/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License