Background: For the efficient and secure running of the power industry, accurate monthly electricity projections are crucial. Due to coupling variations and a variety of data resolutions, current approaches are still unable to accurately extract multidimensional time-series data.
Objective: For monthly electricity consumption forecasting, a multi-time-scale transformation and temporal attention neural network for a temporal convolutional network is proposed.
Methods: First, a multi-time-scale compression model of temporal convolutional network is proposed, which compresses data on several time scales from different resolutions, such as the economy, weather, and historical load. Second, a multi-source temporal attention module is built to further dynamically extract crucial information. Finally, the decoding-encoding and residual connections' structure contributes to the prediction's improved resilience.
Results: The proposed method was compared with the state-of-the-art monthly load forecasting based on two years of historical data in a certain region, demonstrating its effectiveness.
Conclusion: Through the verification of local historical data, the proposed model was contrasted with cutting-edge monthly load forecasting techniques. The obtained results demonstrate the effectiveness.