Enhanced Short-Term Electricity Load Forecasting Using Meteorological Data and KCB-Attention Network

Abstract

Accurate power load forecasting is crucial for building a smart grid. The KCB-Attention model is proposed in this paper to address the limitation of a single model for load forecasting and the complex coupling between seasonal meteorological data. In the feature engineering phase, sequence data reconstruction and alignment are performed for multidimensional weather-load data, and seasonal pattern decomposition is performed by utilising a dynamic time-sliding window strategy. In the forecasting model, by utilising convolutional neural networks (CNN) for efficient feature extraction, bidirectional long-short-term memory (BiLSTM) was introduced to improve the weight calculation of the self-attention mechanism, and Keplerian optimisation algorithm (KOA) adaptively optimises network model hyper-parameters for different seasonal datasets. This multi-scale feature fusion strategy overcomes traditional models' limitations in handling nonlinear interactions between weather variables and load patterns. The results show that the KCB-attention model outperforms existing methods in terms of forecasting accuracy. It has the potential to enhance the reliability of intelligent grid load forecasting.