Short-Term Load Forecasting Method Based on Improved Transformer

Traditional power load forecasting methods often struggle to handle the long-term nonlinear trends in sequence data, which can result in inadequate forecasting accuracy in practical applications, affecting the normal operation of power systems. Therefore, this paper proposes an improved Transformer model to capture the long-term nonlinear trends in power load time series. The model utilizes a Convolutional Neural Network (CNN) in the encoder to capture short-term dependencies within the forecast horizon and integrates a Long Short-Term Memory (LSTM) network with a self-attention mechanism in the decoder to capture long-range dependencies by incorporating periodic components of the power load. This addresses the issue of losing periodic data features due to excessively long load sequences. Additionally, Lagrange interpolation and normalization techniques are used to handle outliers and uneven data distribution in the power load data, ensuring the validity of input data and improving forecasting accuracy at the data level. Experimental results show that, compared to commonly used models like Convolutional Neural Networks-Gate Recurrent Unit (CNN-GRU-Attention) the forecasting error decreases by $\mathbf{3. 4 \% - 5. 3 \%}$.