Short-Term Power Load Forecasting for Industrial Parks Using CNN-BiLSTM Network with Kernel Density Estimation-based Interval Prediction Method

Predicting electrical power loads in industrial parks is challenging due to the uncertainty and variance in model construction. This paper proposes an interval forecasting approach based on a residual-form CNN-BiLSTM, which effectively captures spatial patterns and bidirectional temporal dependencies in time series data. To handle uncertainty, a kernel density estimation (KDE)-based method is employed to generate probabilistic prediction intervals. Experiments using 18 days of real industrial data validate the model’s performance. Results show superior robustness and accuracy compared with LSTM, BiLSTM, and GRU. The proposed model achieves perfect prediction interval coverage probability (PICP = 1.0), narrow normalized interval width (PINAW = 0.0828), and minimal CWC = 0.0828, while baseline models exhibit low coverage or excessively wide intervals. These findings confirm that the method provides both reliable and sharp uncertainty quantification, making it suitable for practical energy forecasting applications.