RESUMO
Due to the crucial role of photovoltaic power prediction in the integration, scheduling and operation of intelligent grid systems, the accuracy of prediction has garnered increasing attention from both the research and industry sectors. Addressing the challenges posed by the nonlinearity and inherent unpredictability of photovoltaic (PV) power generation sequences, this paper introduced a novel PV prediction model known as the dilated causal convolutional network and stacked long short-term memory (DSLSTM). The methodology begins by incorporating physical constraints to mitigate the limitations associated with machine learning algorithms, thereby ensuring that the predictions remain within reasonable bounds. Subsequently, a dilated causal convolutional network is employed to extract salient features from historical PV power generation data. Finally, the model adopts a stacked network structure to effectively enhance the prediction accuracy of the LSTM component. To validate the efficacy of the proposed model, comprehensive experiments were conducted using a real PV power generation dataset. These experiments involved comparing the predictive performance of the DSLSTM model against several popular existing models, including multilayer perceptron (MLP), recurrent neural network (RNN), long short-term memory (LSTM), gated recurrent unit (GRU), stacked LSTM and stacked GRU. Evaluation was performed using four key performance metrics: Mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE) and R-squared (R2). The empirical results demonstrate that the DSLSTM model outperforms other models in terms of both prediction accuracy and stability.
