Medium- and long-term load reliability forecasting method of power system based on CNN-GRU hybrid neural network

• Lagrange interpolation filled monthly gaps in quarterly statistical data. • K-means clustering of normalized features enabled separate models for different clusters. • A CNN-GRU hybrid with attention mechanism was proposed for load forecasting. In medium- and long-term power forecasting research, data collection is intricate and multifaceted; historical samples often contain missing or abnormal values. Incomplete or anomalous data can significantly impair prediction model performance and reduce forecasting accuracy. To address this issue, a method tailored for assessing load reliability over extended time horizons in power systems is proposed. This solution incorporates a structure that merges convolutional and gated recurrent components. Using Lagrange interpolation, gaps in quarterly statistical data are filled on a monthly basis. The normalized input features are clustered via the k-means algorithm, and prediction models are trained separately on datasets with different cluster labels. Considering both static and dynamic characteristics of power load in a multi-frequency combination, an internal model load frequency controller is designed to achieve active disturbance rejection control. By integrating a convolution-based neural processing framework, gated recurrent computational units, and an attention-enhancing module, a model for long-term power system load prediction is proposed. Data analysis demonstrates that the load obtained through the proposed design closely matches the measured values and follows the same variation trend. In terms of performance metrics such as MSE and SMAPE, the proposed method achieves state-of-the-art results. Furthermore, the training process exhibits a faster convergence rate and reduced training time (less than 6 seconds).