1. Introduction
In recent times, the worldwide concentrate on creating clear vitality sources has led to the growing prominence of photo voltaic vitality as a viable different for assembly vitality calls for [1,2]. Photovoltaic (PV) stations are categorized into land-based and water-based stations, relying on their geographical location. Land-based PV stations have a well-established power-generation expertise, whereas water-based PV stations have emerged as a comparatively new kind of facility lately. The PV array is primarily put in on lakes, reservoirs, or oceans, thereby minimizing land utilization [3,4]. This strategy provides the benefit of maximizing the synergistic advantages between energy technology and aquaculture, generally known as the “higher energy technology, decrease aquaculture” idea. Inserting photo voltaic panels on the water’s floor supplies shading for fish, improves the native water temperature, and enhances the financial returns of aquaculture. Moreover, the upper relative humidity across the photo voltaic panels contributes to diminished module temperature and, consequently, enhances the photoelectric conversion effectivity of the photo voltaic panels to some extent. The facility technology of water-based PV stations surpasses that of land-based PV stations when contemplating the identical space [5]. Precisely and swiftly predicting the facility technology of PV stations [6,7,8,9] allows enterprises to develop well timed power-scheduling plans, addressing the problem of temporal mismatch between provide and person demand. This facilitates well timed adjustment of electrical energy costs and maximizes the promotion of person demand for energy consumption, enhancing demand-side flexibility and mitigating extra energy technology. These aims align effectively with China’s “dual-carbon” targets. Consequently, the institution of a high-precision prediction mannequin with robustness and a powerful generalization potential holds vital significance.
With the fast development of synthetic intelligence algorithms, an growing variety of algorithms are being employed for solar energy prediction. Direct prediction using data-driven algorithm fashions has confirmed to be extra dependable and correct in comparison with oblique prediction strategies [10]. Conventional machine-learning methods, resembling principal-component evaluation (PCA) [11,12], variable modulus decomposition (VMD) [13], and random forest (RF) [14], are utilized for data-feature extraction and prediction. Moreover, support-vector machine (SVM) [15], synthetic neural community (ANN) [16], Elman neural community [17], radial foundation operate neural community (RBF) [18], Bayesian strategies [19], amongst others, are employed for power-generation prediction. For example, in references [20,21], SVM was utilized to boost the accuracy of PV power-generation prediction. Reference [22] used VMD to extract function variables and enter them into the ANN mannequin to foretell the output energy of the PV system. In Reference [23], the Okay-means algorithm was utilized to cluster totally different weathers, and Elman served because the prediction mannequin to enhance the prediction accuracy and robustness. The literature [24] employed PCA to extract data-feature data and scale back dimensionality, which was then fed right into a generalized regression neural community (GRNN) mannequin for prediction.
The aforementioned strategies usually exhibit limitations in information mining, as they have a tendency to concentrate on shallow function extraction with out delving into deeper inside function data. Consequently, their data-driven capabilities are comparatively normal [25]. When confronted with the problem of predicting energy technology within the presence of excessive volatility, shallow studying fashions could produce vital errors. To deal with these points successfully, deep studying fashions provide promising options.
The deep studying mannequin displays a extra intricate community construction in comparison with the shallow studying mannequin, enabling it to successfully categorical complicated capabilities and possess deep studying capabilities. Notably, the convolutional neural community (CNN) and time convolutional community (TCN) [26] excel in data-feature extraction. These fashions can adeptly seize temporal dependencies in sequential information, showcasing strong information depth mining and have information-extraction capabilities. Such efficiency will not be attainable with shallow studying fashions. Deep studying fashions, together with stacked autoencoder (SAE) [27,28], deep perception community (DBN) [29,30], recurrent neural community (RNN) [31], and enhanced variants of lengthy short-term reminiscence (LSTM) [32,33] and gated recurrent unit (GRU) [34], provide enhanced accuracy and stability in prediction. In References [1,35], CNN was employed for information filtering and denoising, whereas the pooling layer diminished information dimensionality and reminiscence necessities and improved processing velocity. By establishing deep connections between function data and leveraging its reminiscence unit, LSTM achieved high-precision prediction outputs. References [36,37] proposed a hybrid CNN-GRU mannequin to reduce power-prediction errors. In [38], the accuracy of the LSTM-RNN prediction mannequin was enhanced via a time-dependent correction technique. Reference [39] launched the LSTM-TCN hybrid mannequin, which demonstrated superior prediction efficiency in comparison with single fashions throughout various climate situations and a number of time-prediction intervals.
It’s value noting that the adjustment of hyperparameters is essential for harnessing the sturdy studying potential of deep studying fashions. Handbook adjustment or conventional strategies for hyperparameter optimization may be time-consuming and expensive, making them impractical for enterprises. To effectively tackle this optimization drawback, heuristic optimization algorithms provide a viable answer. These algorithms make use of computerized iteration to quickly converge to the worldwide optimum worth by contemplating constraints, goal capabilities, and resolution variables. They’re able to successfully fixing numerous hyperparameter-optimization mixture issues [31]. Examples of such algorithms embrace genetic algorithm (GA) [40,41,42], particle-swarm optimization (PSO) [43,44], firefly algorithm (FA) [45], ant-colony algorithm (ACO) [21], bat algorithm (BA) [46,47], and others.
Restricted analysis has been carried out on the prediction of fishing–photo voltaic complementary PV stations. On this examine, we suggest a hybrid prediction mannequin, ICMIC-POA-CNN-BIGRU, which leverages historic PV power-generation information and numerical climate prediction (NWP) meteorological information as the unique dataset to drive the algorithm. The CNN part of our mannequin successfully mines the related data between PV energy technology and power-generation variables, extracting deep options from the information. These options are then inputted into the bidirectional gated recurrent unit (BIGRU) part for prediction. BIGRU, consisting of two GRUs, processes sequence information bidirectionally, enabling the mannequin to seize deep function data earlier than and after particular moments and improve its generalization potential. To effectively optimize the hyperparameter mixture of CNN-BIGRU, we make the most of the newest heuristic optimization algorithm, the pelican optimization algorithm (POA). Moreover, the ICMIC chaotic mapping method optimizes the preliminary inhabitants place of POA, stopping the mannequin from getting caught in native optima and bettering international convergence velocity. It’s value noting that the shortcoming of the hybrid mannequin relative to the only mannequin is the rise in computational value. The principle contributions of this examine are summarized as follows:
Combining CNN’s highly effective data-feature extraction potential with BIGRU’s two-way use of time collection prediction, the newest heuristic algorithm (POA) is proposed for the primary time to optimize CNN-BIGRU hyperparameters;
To boost the convergence velocity and international optimization potential of the algorithm, the ICMIC chaotic mapping method is employed to optimize the initialization place of the POA inhabitants;
To deal with potential overfitting points throughout coaching, the L2-regularization method is employed, which facilitates fast convergence of the loss curve for CNN-BIGRU. Moreover, the optimum L2-regularization coefficient is robotically iteratively optimized utilizing the POA algorithm;
The prediction efficiency of the six fashions beneath various climate situations is evaluated utilizing six analysis indicators. Okay-fold cross-validation is carried out to match the prediction results of the three hybrid fashions on totally different datasets spanning three consecutive days and 5 consecutive days. The outcomes display that the hybrid mannequin proposed on this examine displays probably the most superior prediction efficiency.
Part 2 introduces the supply of the information and supplies an in depth rationalization of the information preprocessing strategies and the working rules of the hybrid prediction mannequin; Part 3 discusses the prediction outcomes of the totally different check units; Part 4 attracts a conclusion.
