The application of WUTP-CNN-GRU model in power prediction of desert and floating photovoltaic system Authors: Junxia Jiang, Hongbo Su, Huan Du, Huihui Yin Published: 2025 ABSTRACT Due to the intermittent nature of solar energy and meteorological uncertainty, accurate photovoltaic (PV) power generation prediction is crucial to grid stability. In this study, a new WUTP-CNN-GRU hybrid model is proposed based on the in-depth study and analysis of the measured data of Dongdatan desert and Yangzhong floating PV power stations. Firstly, focusing on the data preprocessing link, the fuzzy C-means clustering (FCM) method is innovatively introduced to effectively divide the weather types, and the prediction models suitable for different working conditions such as sunny, cloudy and rainy days are constructed based on the classification results. The model integrates CNN and GRU architecture, and introduces WUTP algorithm to optimize hyperparameters. The comparison experiments with WOA, PSO and DB0 optimization models show that the estimated performance of this model is significantly better than that of the comparison model, and it shows excellent adaptability and stability under different weather conditions, which provides an effective solution for smart grid management and PV system optimization. Introduction In response to the environmental and energy crisis, the world is actively promoting the transformation of energy systems to renewable energy and is committed to building an efficient, sustainable, low-carbon energy framework1. Faced with the continued growth in energy demand and the depletion of traditional fossil fuels, vigorously developing renewable energy has become a feasible path to alleviate environmental degradation and energy shortages2. Among them, solar energy has been widely used and developed rapidly in recent years due to its clean and pollution-free characteristics and significant economic and environmental advantages3. However, solar PV power generation is significantly affected by weather factors, and its strong randomness and intermittency can easily lead to fluctuations in grid voltage, current and frequency, posing severe challenges to power dispatching4–6. Therefore, the development of accurate PV power generation prediction technology is the key to dealing with its uncertainty7,8. Literature review PV power generation forecast classification is mainly based on the forecast time scale, input data, and forecast method. The following are common forecast methods. According to the prediction time scale, PV power generation forecasts are usually divided into four categories: ultra-short term (0–4 h), short term (1–7 days), medium term (1 week–1 month) and long term (next few months)9. Among them, short-term and ultra-short-term forecasts can provide highly reliable information, which is an important basis for spot transactions in the electricity market and real-time grid dispatching10,11. Current research mainly focuses on the fields of ultra-short-term and short-term forecasts12. However, the inherent strong randomness, intermittency and uncertainty of PV power generation make the screening of training data samples challenging. The accuracy of existing forecasting technologies, especially under unconventional weather conditions such as cloudy and thunderstorms, is still difficult to meet the requirements of practical applications13. According to different prediction mechanisms, PV power generation prediction models are divided into four categories: physical models, statistical models, deep learning models and hybrid models14–16. Physical models rely on accurate system parameters and meteorological data and are suitable for long-term predictions17–19. Statistical models mainly rely on historical data and are good at short-term power prediction20,21. Gao et al.22 proposed to apply discrete gray model (DGM) combined with historical power data of adjacent dates to correct hourly power forecasts for non-ideal weather conditions to overcome the volatility problem of day-ahead predictions. The deep learning model is good at automatically extracting data features through multi-level nonlinear transformation, especially for processing complex data23,24. QPFNet, a deep learning model developed by Zhou et al.25 uses the basic meteorological elements in the ERA5 reanalysis dataset to fit the nonlinear relationship between them and precipitation, and achieves better prediction performance than traditional methods. Sun et al.26 designed a dedicated convolutional neural network ' SUNSET ' to predict the average PV power in the next 15 min. In contrast, the hybrid model shows stronger competitiveness in prediction efficiency and accuracy by integrating the advantages of multiple models. Qu and other27 scholars proposed an attention-based long-term and short-term neural network prediction model (ALSM). The model is assembled using CNN, LSTM and attention mechanism in the multi-correlation and target variable prediction mode to achieve hourly PV power prediction for the next day. It is noteworthy that advanced deep learning approaches have demonstrated remarkable performance in time-series prediction tasks across various domains. For instance, in trajectory prediction, the CORR-CNN-BiLSTM-Attention model28 and temporal convolutional network with dual attention mechanisms29 have shown excellent capability in capturing spatiotemporal dependencies, providing valuable insights for PV power forecasting that similarly involves complex temporal patterns. Furthermore, optimization algorithms for hyperparameter tuning have been extensively employed in engineering applications. Improved optimization techniques such as the Enhanced Dung Beetle Optimizer30 and Harris Hawk Optimization31 have demonstrated superior performance in parameter identification and prediction tasks, while Extreme Learning Machines32 have shown efficiency in regression problems. These methodologies offer promising alternatives for enhancing PV prediction model performance. In the broader context of power system forecasting, deep learning approaches have also shown significant advances. For probabilistic load forecasting, the non-crossing sparse group Lasso-quantile regression deep neural network33 represents a state-of-the-art approach that could be adapted for probabilistic PV power forecasting. For floating systems specifically, the challenges of distributed data learning in marine environments have been addressed through innovative approaches like event-triggered federated learning for offshore wind turbine fault diagnosis34, which could provide valuable insights for distributed floating PV system monitoring and prediction. Research deficiencies and contributions The above research shows that due to the intermittent and uncertain nature of solar energy, PV data is easily affected by weather conditions. The classification and prediction of different weather types is another difficult point in PV power prediction, especially the prediction of different weather types, whose accuracy will directly affect the effectiveness of classification forecast35. Despite these advances, several critical research gaps remain in PV power prediction: (1) Training Data Limitations Due to the high randomness of solar energy, the correlation between data samples is low. Most existing studies choose historical data as training sets, which contain many irrelevant or low-correlation samples, adversely affecting model robustness. (2) Model Adaptability Challenges For unconventional weather conditions, insufficient historical data makes it difficult for models to learn the complex relationship between weather variables and PV output. Although separate CNN and GRU models have been widely studied, research on their integration for PV prediction remains limited. (3) Limited Cross-Domain Methodology Transfer While advanced deep learning and optimization approaches have proven effective in related domains (trajectory prediction, parameter identification, load forecasting), their application to PV power prediction, particularly for specialized scenarios like floating systems, remains underexplored. (4) Inadequate Handling of Complex Environments Floating PV systems present unique challenges due to their aquatic environment, including water surface reflectance, humidity effects, and wave-induced oscillations, which are not adequately addressed in current prediction models. To address these limitations, this study makes the following contributions: ①We propose a novel WUTP-CNN-GRU hybrid model that integrates plant-inspired optimization with deep learning architecture for enhanced PV power prediction. ②We introduce a weather-type classification framework using fuzzy C-means clustering to improve model adaptability across different meteorological conditions. ③We conduct comprehensive experiments on both desert and floating PV systems, demonstrating the model’s robustness across diverse environmental conditions. ④We establish extensive benchmarks including classical statistical models (ARIMA, SVR), modern deep learning architectures (Transformer), and optimization algorithms (WOA, PSO, DB0) to thoroughly validate our approach. Article organization The structure of this paper is organized to facilitate readers’ comprehension and evaluation. The remaining chapters are organized as follows: Section "Methods" details the data source and preprocessing, model composition, optimization algorithm, and accuracy evaluation method. Section "Case analysis" presents the experimental results, analyzes model performance, and provides comprehensive discussion. Section "Conclusion" summarizes the main conclusions and suggests future research directions. Methods In the actual scenario of PV power prediction, due to mechanical failures of equipment, extreme weather disturbances, and human operational errors, the original data generally has problems such as missing records and numerical anomalies, which greatly affects the accuracy of the prediction36. The main purpose of this study is to propose a robust and high-precision PV power prediction method using a hybrid model. The following sections detail the source and preprocessing of the data, the composition of the hybrid model, the optimization algorithm, and the accuracy evaluation method. Data source and preprocessing To systematically evaluate the adaptability and robustness of the proposed model, this study aims to rigorously verify its cross-scenario generalization capability by comparing performance under extreme environmental conditions. For this purpose, we deliberately selected two PV power stations with sharply contrasting climatic and geographical conditions as test cases: the Dongdatan Desert PV Station in Gansu Province, representing a typical continental arid climate, and the Yangzhong Floating PV Station in Jiangsu Province, representing a humid subtropical climate. The primary objective of this site selection strategy is not to achieve extensive geographical coverage, but to construct an effective stress test bed for examining the model’s generalization capability. The data in this article comes from the Dongdatan PV Power Generation Project. The region has a typical continental temperate arid climate, with sufficient sunlight, dry climate, large temperature differences between day and night and four seasons, and longer daylight hours in summer. The dataset includes 8 key meteorological parameters, including TSR, AT, PT, WS, WD, P, RH, and RP, totaling 9761 sets of data. In order to eliminate the influence of inter-feature dimensions and data comparability, the data samples are normalized, and the formula is shown in Eq. (1). After the model is trained, Eq. (2) is used for denormalization.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}^{*}=\frac{{x}_{i}-{x}_{min}}{{x}_{max}-{x}_{min}}$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}={x}_{i}^{*}\left({x}_{max}-{x}_{min}\right)+{x}_{min}$$\end{document} The Spearman correlation coefficient method is used to analyze meteorological factors. The main factors affecting PV power generation are selected and their correlation with meteorological factors such as TSR, AT, PT, WS, WD, P, RH, and RP is analyzed. The analysis results show that solar radiation parameters (TSR) show the strongest correlation with PV output power. Based on this, the meteorological variables selected for model input are TSR, AT, PT WS and RP. The remaining parameters (WD, P, RH) are excluded from the model input because they have a low correlation with PV output power and are considered redundant features37. Fuzzy C-means clustering algorithm FCM is a classic clustering algorithm and is currently the most widely used clustering algorithm38,39. By introducing the concept of “membership”, data points are allowed to belong to multiple clusters with different probabilities. FCM is more suitable for data with fuzzy boundaries (such as weather state transitions and mixed meteorological characteristics). For each data point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}$$\end{document} and cluster j, the membership \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u}_{ij}\epsilon [\text{0,1}]$$\end{document} is defined to satisfy Eq. (3)3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sum_{j=1}^{c}{u}_{ij}=1 \forall j=\text{1,2},3,\cdots ,n$$\end{document} The formula for minimizing the weighted distance square sum is shown in formula (4):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J\left(U,V\right)=\sum_{i=1}^{n}\sum_{j=1}^{c}{u}_{ij}^{m}\cdot {d}^{2}\left({x}_{i},{v}_{j}\right)$$\end{document} The iterative optimization steps are as follows: ①Initialization: Randomly generate a membership matrix U whose row sum is 1. ②Update cluster center: The calculation formula is shown in (5).5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{j}=\frac{\sum_{i=1}^{n}{u}_{ij}^{m}\cdot {x}_{i}}{\sum_{i=1}^{n}{u}_{ij}^{m}}$$\end{document} ③Update membership: The calculation formula is shown in formula (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u}_{ij}=\frac{1}{\sum_{k=1}^{c}{\left(\frac{d\left({x}_{i},{v}_{j}\right)}{d\left({x}_{i},{v}_{k}\right)}\right)}^{\frac{2}{m-1}}}$$\end{document} ④Termination condition: When \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u}_{ij}$$\end{document} satisfies formula (7), the iteration process stops.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$max\left\{\left|{u}_{ij}^{\left(q+1\right)}-{u}_{ij}^{\left(q\right)}\right|\right\}<\varepsilon$$\end{document} The calculation process of the FCM clustering algorithm is shown in Fig. 1.Fig. 1The calculation process of the FCM clustering algorithm. The calculation process of the FCM clustering algorithm. Model composition and algorithm principle Long short-term memory (LSTM) and gated recurrent neural network (GRU) LSTM is an enhanced recurrent neural network designed specifically for sequence modeling. Its core structure includes memory cells and three gating mechanisms (input gate, forget gate, and output gate), which work together to dynamically capture and transmit long-term dependencies40,41. GRU, as a new architecture with a similar concept to LSTM, also optimizes sequence modeling capabilities through gating mechanisms. Its streamlined design uses reset gates and update gates to replace the three-gate structure of LSTM, reducing model complexity while retaining the ability to capture long-term dependencies42,43. Water uptake and transport of lants (WUTP) WUTP is a novel meta-heuristic optimization algorithm, which is inspired by the water transport mechanism and transmembrane flow dynamics in plants44. The algorithm simulates the complete process of water absorption from soil to roots and transport from xylem to leaves, and draws on the adaptive characteristics of plant water management to efficiently deal with high-dimensional continuous and mixed optimization problems. The core mechanism of WUTP is to simulate the change of pressure gradient in water flow, so as to balance the global exploration and local utilization behavior. Its mathematical model provides a theoretical guarantee for efficient search in solution space. Different from PSO, WOA, DB0 and other algorithms based on animal behavior, WUTP algorithm is unique, and its design inspiration comes from the water transport mechanism in plants. The algorithm uses pressure gradient as the core power to simulate the physical and physiological processes of water transport from high water potential area to low water potential area in plants. Its biggest advantage lies in the pioneering introduction of plant bionic principle, and with the help of this mechanism, it achieves a highly adaptive balance between exploration and development, showing unique optimization performance. As a new meta-heuristic optimization algorithm, the optimization performance of WUTP is directly affected by the setting of core parameters such as population initialization and water flow movement. The specific parameters used in this study are detailed in Table 1.Table 1The core parameter setting of WUTP optimization algorithm.Parameter type checkParameter nameValuePopulation initialization parametersno_particles30Flow motion parametersT100K1 × 10−9D1 × 10−9Lp1 × 10−9rho1000g9.81State-switch parameterp0.5Convergence adjustment parametersnu0.0018 Table 1The core parameter setting of WUTP optimization algorithm.Parameter type checkParameter nameValuePopulation initialization parametersno_particles30Flow motion parametersT100K1 × 10−9D1 × 10−9Lp1 × 10−9rho1000g9.81State-switch parameterp0.5Convergence adjustment parametersnu0.0018 The core parameter setting of WUTP optimization algorithm. As shown in Fig. 2 based on the efficient feature extraction capability of CNN and the performance advantage of GRU in time series input data mining. This paper constructs a PV power prediction model based on WUTP-CNN-GRU, which is a prediction algorithm based on WUTP, LSTM and CNN-GRU. Compared with other prediction models, this model has higher prediction accuracy and faster training convergence speed. WUTP is used to find the optimal hyperparameters of the CNN-GRU prediction model and optimize the objective function through continuous optimization. WUTP significantly improves the prediction performance of the CNN-GRU model.Fig. 2CNN-GRU prediction algorithm based on WUTP optimization. CNN-GRU prediction algorithm based on WUTP optimization. Statistical methods Three widely used evaluation indicators, namely root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R2), were used to evaluate the prediction performance of each model45–47. These evaluation indicators enable us to measure the prediction ability of the model from different perspectives. The formulas are as follows (8)–(10):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAE=\frac{1}{N}\sum_{i=1}^{N}|{P}_{i}-{\widehat{P}}_{i}|$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$RMSE=\sqrt{\frac{{\sum }_{i=1}^{N}{\left({P}_{i}-{\widehat{P}}_{i}\right)}^{2}}{N}}$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}^{2}=1-\frac{{\sum }_{i=1}^{N}{\left({P}_{i}-{\widehat{P}}_{i}\right)}^{2}}{{\sum }_{i=1}^{N}{\left({\overline{P} }_{i}-{P}_{i}\right)}^{2}}$$\end{document} This study proposes a systematic hybrid modeling framework, the overall workflow of which is illustrated in Fig. 3 and comprises the following three key phases:Fig. 3The overall framework of WUTP-CNN-GRU hybrid model. The overall framework of WUTP-CNN-GRU hybrid model. First, in the data preprocessing and weather pattern classification stage, raw data undergo cleaning and normalization. Subsequently, fuzzy C-means clustering is employed to categorize weather conditions into several typical types (sunny, cloudy, and rainy) based on key meteorological features, thereby providing a structured input foundation for subsequent modeling. Second, in the parallel sub-model construction and optimization phase, an independent CNN-GRU prediction sub-model is built for each specific weather type. This model first extracts spatial-local features from the input sequences via CNN, and then captures long-term temporal dependencies using GRU. To further enhance model performance, the WUTP optimization algorithm is introduced as an automated hyperparameter tuning mechanism, forming a closed-loop optimization process guided by prediction errors such as RMSE. By continuously searching for key hyperparameters (including learning rate, number of layers, and quantity of convolutional filters, among others), the algorithm significantly improves the prediction accuracy and training convergence efficiency of each sub-model. Finally, in the prediction output stage, the system automatically identifies the weather category of real-time meteorological data and invokes the corresponding WUTP-optimized CNN-GRU sub-model to generate high-accuracy power prediction results. Case analysis On the same PV power generation dataset, the performance of different prediction models shows significant differences. In order to comprehensively evaluate the prediction accuracy of the CNN-GRU model proposed in this paper and its adaptability in a changing environment, this study systematically compared it with a variety of representative time series prediction models, including LSTM, GRU, CNN-LSTM and CNN-BILSTM. The above models have been applied to PV power generation prediction tasks under a variety of typical weather conditions. Figure 3 intuitively shows the comparison between the prediction results of each model under different weather conditions and the actual power generation curve. As shown in Fig. 4 in the PV power prediction task under different weather conditions, the prediction curve of the CNN-GRU model is highly consistent with the actual observation value, and its trajectory is closely consistent and the fluctuation characteristics are highly consistent. In contrast, the prediction results of other basic models show different degrees of deviation. This visual comparison fully proves that the proposed model has excellent environmental robustness and time series prediction accuracy, and can effectively capture the nonlinear dynamic characteristics of PV systems under complex meteorological conditions. In order to further quantify the performance differences of each model, Table 1 systematically summarizes the comparison results of key evaluation indicators (including MAE, RMSE, R2 and MSE), and provides data support for model advantages.Fig. 4Prediction results under different weather conditions. (a) Sunny day (b) Overcast day (c) Rainy day. Prediction results under different weather conditions. (a) Sunny day (b) Overcast day (c) Rainy day. As shown in Table 2, the evaluation indicators of the CNN-GRU hybrid model in this study highlight significant competitive advantages. Compared with the traditional LSTM model and other comparative benchmarks, CNN-GRU has shown better prediction performance under various complex meteorological conditions. Not only has it achieved higher prediction accuracy on this dataset, but more importantly, it has shown excellent robustness and generalization adaptability when dealing with PV power prediction tasks under different weather conditions (such as sunny, overcast, rainy). It has the lowest error and the highest determination coefficient, indicating its superiority in prediction tasks.Table 2Comparison of model evaluation structures under different weather conditions.ClimaticPredictionAssessment criteriaRMSEMAEMSER2ARIMASVR7.33186.78135.39644.444253.755345.98560.90590.92782Sunny dayLSTM6.91344.768347.79450.9247GRU5.55834.123730.89510.9832CNN-LSTM6.38914.472540.82080.9305CNN-BILSTMTransformer6.53875.49794.68513.729742.755230.22640.91680.9869CNN-GRUARIMASVR5.05807.22386.64263.43134.88654.876823.735152.182844.12400.99490.90410.9213Overcast dayLSTM6.77484.507145.89730.9244GRU5.94234.373038.71710.9417CNN-LSTM6.36914.590240.56560.9300CNN-BILSTMTransformer5.65655.32723.68063.308736.309235.13120.94160.9640CNN-GRUARIMASVR5.08987.30926.81562.58775.26214.756726.607953.423846.45180.99770.90430.9224Rainy dayLSTM6.79654.821046.19180.9234GRU5.52954.078930.57510.9561CNN-LSTM6.40074.442335.96940.9657CNN-BILSTMTransformer6.61365.95164.47903.991938.739335.3217096070.9639CNN-GRU5.08713.552827.82060.9961 Table 2Comparison of model evaluation structures under different weather conditions.ClimaticPredictionAssessment criteriaRMSEMAEMSER2ARIMASVR7.33186.78135.39644.444253.755345.98560.90590.92782Sunny dayLSTM6.91344.768347.79450.9247GRU5.55834.123730.89510.9832CNN-LSTM6.38914.472540.82080.9305CNN-BILSTMTransformer6.53875.49794.68513.729742.755230.22640.91680.9869CNN-GRUARIMASVR5.05807.22386.64263.43134.88654.876823.735152.182844.12400.99490.90410.9213Overcast dayLSTM6.77484.507145.89730.9244GRU5.94234.373038.71710.9417CNN-LSTM6.36914.590240.56560.9300CNN-BILSTMTransformer5.65655.32723.68063.308736.309235.13120.94160.9640CNN-GRUARIMASVR5.08987.30926.81562.58775.26214.756726.607953.423846.45180.99770.90430.9224Rainy dayLSTM6.79654.821046.19180.9234GRU5.52954.078930.57510.9561CNN-LSTM6.40074.442335.96940.9657CNN-BILSTMTransformer6.61365.95164.47903.991938.739335.3217096070.9639CNN-GRU5.08713.552827.82060.9961 Comparison of model evaluation structures under different weather conditions. ARIMA SVR 7.3318 6.7813 5.3964 4.4442 53.7553 45.9856 0.9059 0.9278 2 Sunny day CNN-BILSTM Transformer 6.5387 5.4979 4.6851 3.7297 42.7552 30.2264 0.9168 0.9869 CNN-GRU ARIMA SVR 5.0580 7.2238 6.6426 3.4313 4.8865 4.8768 23.7351 52.1828 44.1240 0.9949 0.9041 0.9213 CNN-BILSTM Transformer 5.6565 5.3272 3.6806 3.3087 36.3092 35.1312 0.9416 0.9640 CNN-GRU ARIMA SVR 5.0898 7.3092 6.8156 2.5877 5.2621 4.7567 26.6079 53.4238 46.4518 0.9977 0.9043 0.9224 CNN-BILSTM Transformer 6.6136 5.9516 4.4790 3.9919 38.7393 35.3217 09607 0.9639 In order to further demonstrate the superiority of the CNN-GRU model, the WOA, PSO, DB0, and WUTP algorithms are used to optimize the CNN-GRU. The PV power prediction results of each model under rainy conditions are shown in Fig. 5. It can be seen from the figure that the prediction curve of the WUTP-CNN-GRU model is consistent with the actual curve. It can be seen from Table 3 that the results of the WUTP-CNN-GRU model are better. In terms of R2, it is 2.59%, 1.59%, and 0.99% higher than the WOA model, PSO model, and DB0 model, respectively. RMSE decreased by 0.3061, 0.214, and 0.1473, respectively. MAE decreased by 0.1962, 0.1334, and 0.0905, respectively. MSE decreased by 0.2349, 0.1445, and 0.0896, respectively.Fig. 5Prediction results of different algorithm optimization models.Table 3Results of different algorithm optimization models.ModelRMSEMAEMSER2WOA-CNN-GRU0.53660.34070.28800.9683PSO-CNN-GRU0.44450.27790.19760.9783DB0-CNN-GRU0.37780.23500.14270.9843WUTP-CNN-GRU0.23050.14450.05310.9942 Prediction results of different algorithm optimization models. Table 3Results of different algorithm optimization models.ModelRMSEMAEMSER2WOA-CNN-GRU0.53660.34070.28800.9683PSO-CNN-GRU0.44450.27790.19760.9783DB0-CNN-GRU0.37780.23500.14270.9843WUTP-CNN-GRU0.23050.14450.05310.9942 Results of different algorithm optimization models. In order to improve the prediction accuracy of the verification model, the 21-year data in Jiangsu Yangzhong station were selected for prediction. WD, WS, P, T and DSR were used as the input features of the model, and RP was used as the output feature. The weather was classified and optimized by optimization algorithm. The Fig. 6 shows the prediction results under three weather types on the lake surface. In addition, Table 4 gives the evaluation indexes of different models. It can be seen from the table that the performance of CNN-GRU model is the best under any working condition, and its R2 value is above 0.99. It can be seen that CNN extracting spatial features can effectively improve the accuracy and robustness of time series prediction.Fig. 6Prediction results under different weather conditions in Yangzhong station. (a) Sunny day (b) Overcast day (c) Rainy day.Table 4Comparison of model evaluation structures under different weather conditions in Yangzhong station.ClimaticPredictionAssessment criteriaRMSEMAEMSER2ARIMASVR6.06565.89475.34834.873636.791834.74740.97340.9789Sunny dayLSTM5.68134.232132.27690.9799GRU5.47524.867929.97790.9753CNN-LSTM5.16714.332926.69910.9773CNN-BILSTMTransformer5.37624.81063.82403.633428.904123.14190.97760.9874CNN-GRUARIMASVR4.33576.05235.80512.69804.38414.417218.798136.630933.69930.99740.97270.9737Overcast dayLSTM6.09725.158236.95660.9724GRU5.38764.816929.02630.9622CNN-LSTM5.63264.418031.72660.9728CNN-BILSTMTransformer5.05124.78523.68213.308725.514922.89810.97580.9870CNN-GRUARIMASVR4.30946.44115.97652.87334.95084.328818.570741.488035.71880.99700.94500.9664Rainy dayLSTM6.03444.294736.40220.9550GRU5.83134.398734.00400.9564CNN-LSTM5.52384.116430.51250.9635CNN-BILSTMTransformer5.10964.02493.86102.523926.108516.1996097290.9879CNN-GRU3.80792.176114.50010.9981 Prediction results under different weather conditions in Yangzhong station. (a) Sunny day (b) Overcast day (c) Rainy day. Table 4Comparison of model evaluation structures under different weather conditions in Yangzhong station.ClimaticPredictionAssessment criteriaRMSEMAEMSER2ARIMASVR6.06565.89475.34834.873636.791834.74740.97340.9789Sunny dayLSTM5.68134.232132.27690.9799GRU5.47524.867929.97790.9753CNN-LSTM5.16714.332926.69910.9773CNN-BILSTMTransformer5.37624.81063.82403.633428.904123.14190.97760.9874CNN-GRUARIMASVR4.33576.05235.80512.69804.38414.417218.798136.630933.69930.99740.97270.9737Overcast dayLSTM6.09725.158236.95660.9724GRU5.38764.816929.02630.9622CNN-LSTM5.63264.418031.72660.9728CNN-BILSTMTransformer5.05124.78523.68213.308725.514922.89810.97580.9870CNN-GRUARIMASVR4.30946.44115.97652.87334.95084.328818.570741.488035.71880.99700.94500.9664Rainy dayLSTM6.03444.294736.40220.9550GRU5.83134.398734.00400.9564CNN-LSTM5.52384.116430.51250.9635CNN-BILSTMTransformer5.10964.02493.86102.523926.108516.1996097290.9879CNN-GRU3.80792.176114.50010.9981 Comparison of model evaluation structures under different weather conditions in Yangzhong station. ARIMA SVR 6.0656 5.8947 5.3483 4.8736 36.7918 34.7474 0.9734 0.9789 CNN-BILSTM Transformer 5.3762 4.8106 3.8240 3.6334 28.9041 23.1419 0.9776 0.9874 CNN-GRU ARIMA SVR 4.3357 6.0523 5.8051 2.6980 4.3841 4.4172 18.7981 36.6309 33.6993 0.9974 0.9727 0.9737 CNN-BILSTM Transformer 5.0512 4.7852 3.6821 3.3087 25.5149 22.8981 0.9758 0.9870 CNN-GRU ARIMA SVR 4.3094 6.4411 5.9765 2.8733 4.9508 4.3288 18.5707 41.4880 35.7188 0.9970 0.9450 0.9664 CNN-BILSTM Transformer 5.1096 4.0249 3.8610 2.5239 26.1085 16.1996 09729 0.9879 As shown in Fig. 7, the comparison of the prediction results of CNN-GRU models under different optimization algorithms shows that the WUTP-CNN-GRU model can accurately capture the fluctuation characteristics of the actual power curve and show the best fitting performance in all evaluation models. Table 5 lists the evaluation indexes of the corresponding models of four optimization algorithms (WOA, PSO, DB0, WUTP) in detail. Among them, the R2 values of the models based on WOA, PSO, DB0 and WUTP are 0.9607, 0.9658, 0.9759 and 0.9926, respectively. It is worth emphasizing that the highest R2 value (0.9926) obtained by the WUTP model fully highlights the excellent predictive ability and stability of the optimization model.Fig. 7The results of CNN-GRU model optimized by different algorithms in Yangzhong station.Table 5Evaluation of CNN-GRU optimized by different algorithms in Yangzhong station.ModelRMSEMAEMBER2WOA-CNN-GRU236.3904161.704924.53270.9607PSO-CNN-GRU220.3038145.911317.23050.9658DB0-CNN-GRU185.1767127.32402.09240.9759WUTP-CNN-GRU102.811371.03028.63150.9926 The results of CNN-GRU model optimized by different algorithms in Yangzhong station. Table 5Evaluation of CNN-GRU optimized by different algorithms in Yangzhong station.ModelRMSEMAEMBER2WOA-CNN-GRU236.3904161.704924.53270.9607PSO-CNN-GRU220.3038145.911317.23050.9658DB0-CNN-GRU185.1767127.32402.09240.9759WUTP-CNN-GRU102.811371.03028.63150.9926 Evaluation of CNN-GRU optimized by different algorithms in Yangzhong station. Conclusion This study innovatively proposes a novel WUTP-CNN-GRU model specifically designed for PV power prediction, which deeply integrates advanced technologies including CNN, GRU, and WUTP. Significant innovations are demonstrated in the following aspects: Algorithm Integration Innovation A WUTP-optimized CNN-GRU hybrid model is proposed for power prediction in both desert and floating PV systems. The plant-inspired WUTP algorithm is introduced to PV power prediction for the first time, achieving deep integration with the CNN-GRU architecture and enabling adaptive hyperparameter optimization that overcomes the tendency of conventional optimization algorithms to converge to local optima. In comparative experiments with other benchmark models, the WUTP-CNN-GRU model consistently demonstrates superior performance, achieving higher R2 values while maintaining the lowest MAE, MSE, and RMSE metrics. Structural Design Innovation Key features are effectively screened through Spearman correlation analysis to eliminate redundancies. A parallel "weather clustering + sub-model" prediction framework is implemented, where FCM clustering enables precise weather type classification, and dedicated CNN-GRU sub-models are constructed for distinct weather conditions. This design significantly enhances the model’s adaptability to variable meteorological patterns, demonstrating excellent robustness. Significant Performance Improvement Comprehensive validation conducted across two typical climatic regions—the arid Dongdatan and the humid Yangzhong Station—shows that the model not only significantly outperforms a series of baseline models in prediction accuracy but also exhibits exceptional cross-scenario adaptability. This provides a more universal solution for power prediction in PV systems across diverse geographical locations. Although the proposed WUTP-CNN-GRU model demonstrates excellent performance under typical meteorological conditions, it still exhibits certain limitations. The current model has not been validated for predicting extreme weather events such as typhoons and sandstorms. Enhancing predictive capability for extreme weather requires integrating multi-source data including satellite cloud images, radar, and NWP to develop a more perceptive forecasting framework. Additionally, the training process incurs high computational costs due to hyperparameter search. Improving engineering application efficiency involves developing distributed and parallelized versions of the WUTP algorithm to facilitate rapid deployment and application of the model in ultra-large-scale PV clusters.