An Adaptive Variable-Weight Combination Forecasting Method for Energy Product Sales Based on Meta-Learning

Accurate energy product sales forecasting is one of the core tasks of energy planning modeling and is crucial for the sustainability of the energy supply chain. However, energy product sales data have both linear and nonlinear features, and it is difficult for a single model to effectively balance the feature learning of both. In addition, variable influence of external environment leads to the variability of each feature, and traditional fixed-weight combination multi-step forecasting is difficult to adapt to such variability. To address these problems, this study proposes a dynamic combination forecasting method that can adaptively change the model combination weights according to external environment changes. Firstly, a comprehensive and scientific feature system is constructed by combining natural language processing and statistical measurement method. Secondly, Lasso and LSTM models are constructed to deal with overlapping linear and nonlinear features in sales data, respectively. Finally, the meta-learning method is used to establish a relationship function between model combination weights and external features, which realizes adaptive variable model combination for different moments and energy products. Experimental results based on real data from 850 gasoline stations in China show that the proposed method (External feature-based variable-weight combination forecasting model, EFVCFM) significantly outperforms the baseline models and other combination models in terms of forecasting accuracy and stability, with a MAPE and R2 of 5.04% and 0.92, respectively. In addition, the method has high robustness and still achieves excellent model combination results on new data.