| Popis: |
Precisely forecasting output of solar photovoltaics is crucial for (i) effective solar power management, (ii) integration into the electrical grid, (iii) flexible allocation of power resources. While deep learning algorithms have shown promise in energy applications, single algorithms often struggle with unstable predictions and limited generalizability for predicting photovoltaic (PV) output. This study introduces an innovative hybrid model (HWGC-WPD-LSTM) that integrates an improved similar day algorithm (WGC: weighted grey correlation analysis and cosine similarity), Wavelet Packet Decomposition (WPD), and Long Short-Term Memory neural network (LSTM) for predicting day-ahead power output. The model suggests an approach to identifying similar days by integrating weighted GRA with cosine similarity. It then decomposes power sequences employing WPD to capture various frequency characteristics. Four independent LSTM networks are then applied to these sub-sequences to forecast output, which are then reconstructed to derive the ultimate forecast outcome for solar photovoltaics. The evaluation of the hybrid model is conducted based on data gathered from actual generating station in Shandong Province, China. Then it is compared against other models utilizing similar day selection methods and other hybrid HWGC-BP, HWGC-Elman, HWGC-SVM, HWGC-RF, and HWGC-LSTM models. This comparison is based on four performance metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), normalized Root Mean Square Error (NRMSE), and Mean Absolute Deviation (MAD). Results demonstrate that the HWGC-WPD-LSTM model offers enhanced precision and stability (MAE = 0.2168 MW, RMSE = 0.2996 MW, NRMSE = 6.78 %, MAD = 2.18 %) in day-ahead power generation predictions. This highlights the potency of the hybrid model in enhancing the forecasting capabilities for solar photovoltaics, which is crucial for the strategic enhancement of renewable energy resource exploitation in the context of modern power systems. |
