Explainable Hybrid Deep Learning Model with Attention Mechanism for Short-Term Load Forecasting

Effective energy forecasting is essential for optimizing electricity management, enhancing efficiency, and reducing operational costs. Traditional forecasting models often struggle to capture complex consumption patterns, limiting their reliability in real-world applications. This study introduces a hybrid deep learning model that integrates Hybrid Dilated Convolutional Neural Networks (HyDCNN) and Long Short-Term Networks (LSTNet), enhanced with a multi-head attention mechanism inspired by the Temporal Pattern Attention LSTM (TPA-LSTM). The model is designed to improve both the accuracy and interpretability of short-term load forecasting. The model was evaluated on two real-world datasets from Grenoble and New South Wales. Experimental results show that it outperforms existing deep learning methods by up to 10%, demonstrating improved forecasting accuracy over state-of-the-art models. Additionally, compared to traditional statistical models, the hybrid approach achieves improvements of up to 79.80%, highlighting its superior capability to model complex energy consumption patterns. To enhance transparency, Integrated Gradients was applied, allowing for a deeper understanding of how specific time steps and input features influence predictions. This research contributes to more reliable and explainable energy forecasting by integrating Hybrid Dilated Convolutional Neural Networks (HyDCNN) and Long Short-Term Networks (LSTNet), which together improve both forecasting accuracy and interpretability. This hybrid approach allows for better handling of short-term fluctuations and long-term dependencies, making it particularly suitable for smarter energy management strategies in applications such as smart buildings and real-time grid optimization. The novelty lies in combining high-accuracy deep learning techniques with interpretability tools, ensuring energy predictions are not only precise but also transparent for real-world decision-making. Furthermore, insights gained from this study can aid in the development of scalable forecasting models applicable to diverse energy management scenarios, including grid stability, demand-side response, and operational optimization in electricity markets.