Abstract
Electrical load forecasting is essential for the efficient operation and planning of power systems. Recent studies have employed Transformer models in forecasting due to their unique attention mechanisms and ability to extract correlations in data. However, these models face challenges in integrating varied data types and capturing long-term dependencies. To address these limitations, this study proposes a TFTformer, a transformer-based neural network designed to enhance the accuracy and generalisability of load forecasting models. The TFTformer incorporates transposed feature-specific embeddings for weather, time, and load data to more accurately capture their unique characteristics. A linear transformation layer post embedding improves feature representation, aligning and standardising features across sequences for improved pattern recognition. Additionally, a Temporal Convolutional Network is integrated within the Transformer’s encoder, employing causal convolutions and dilation to adapt to the sequential nature of data with an expanded receptive field. The effectiveness of the TFTformer is demonstrated through a comparative study against several state-of-the-art methods using load datasets from Belgium, New Zealand, and five Australian states. The results demonstrate that the TFTformer achieves significant MSE improvements across different locations, with over 50% improvement over most models, 42% over CARD, and 16%–17% improvement compared to iFlowformer and iReformer. Furthermore, an Analysis of Variance is conducted to evaluate the impact of each component of the TFTformer. A SHAP-based interpretability analysis, using surrogate models, is conducted to elucidate the decision-making process of TFTformer, highlighting the critical role of time factors and weather features in its predictions.
• Introducing transposed feature-specific embeddings for load, weather, and time data.
• Linear transformation post-embedding standardises features for better recognition.
• Temporal Convolutional Network enhances TFTformer to capture long-term dependencies.
• TFTformer shows superior forecasting accuracy across multiple datasets and regions.
• An ANOVA-based ablation study highlights the contribution of each model component.