Abstract
Accurate day-ahead power generation forecasting is crucial for improving the operational efficiency of energy storage power stations and enhancing the reliability of power grid dispatch. To address the challenges of prediction inaccuracy stemming from the complex nature of energy storage power stations, the difficulties in quantifying charge-discharge energy losses, and the obstacles in extracting implicit features from historical power data, this paper proposes a hybrid forecasting method that integrates chaos theory, signal decomposition, and deep learning, optimized using an adaptive genetic algorithm. First, a refined loss model is established for the battery packs, power conversion systems (PCS), transformers, and station auxiliary power consumption, to quantify energy losses during operation. Second, chaotic phase-space reconstruction is applied to the historical power generation data to reveal its inherent dynamic characteristics. Subsequently, the reconstructed sequences are processed using the Ensemble Empirical Mode Decomposition (EEMD) method to obtain a series of Intrinsic Mode Function (IMF) components. Based on these components, a Peak-based Frequency Band Division (PFBD) method is employed to aggregate the IMFs into high-frequency and low-frequency feature components, thereby effectively extracting implicit information from the original power sequences. Subsequently, a hybrid forecasting model integrating a Convolutional Neural Network (CNN), a Long Short-Term Memory Network (LSTM), and a Multi-Layer Perceptron (MLP) - denoted as CNN-LSTM-MLP - is constructed. This model takes as its input a combination of the extracted implicit features and the outputs from the loss model. The CNN captures local spatial patterns, the LSTM learns long-term temporal dependencies, and the MLP performs feature integration and nonlinear mapping. Finally, an Adaptive Genetic Algorithm (AGA) is used to automatically optimize the hyperparameters of the hybrid model, thereby improving forecasting performance. Experiments using actual operational data from a 10 MW/20 MWh electrochemical energy storage power station demonstrate that the proposed method achieves excellent performance in day-ahead 24-hour forecasting. The Mean Squared Error (MSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²) reach 6.41 MW², 7.68 MW, and 0.898, respectively. Both forecasting accuracy and goodness of fit significantly outperform those of the compared single or hybrid models. This study provides an effective solution that integrates data-driven approaches with physical models for accurate power forecasting in energy storage power stations.