A transformer-based deep learning model for predicting residential load demand

Abstract

In this dissertation, a transformer-based deep learning model is introduced to solve the problem of predicting household energy consumption time series data. As the demand for high-precision energy consumption prediction in smart energy management systems increases, traditional methods still have limitations in capturing complex dependencies and handling large-scale datasets. The proposed model harnesses the unique capabilities of the Transformer architecture, particularly its attention mechanism, to adeptly capture the long-term dependencies prevalent in energy consumption data, thereby enhancing the accuracy of load demand for residential-level users. The project optimized the Transformer encoder layer to support batch data processing and improve training efficiency. MSE loss function and Adam optimizer are employed to ensure fast and stable convergence. Through comparative experiments with traditional LSTM models, this model shows significant advantages in terms of accuracy, prediction speed, and ability to handle large-scale datasets, demonstrating the strong potential of utilizing deep learning techniques to process complex data in time series. The results of the model on residential energy consumption prediction task demonstrate its potential application value in future smart energy management systems, and future work is expected to explore the application of the model in different energy management scenarios, as well as its contribution to promoting sustainable energy utilization.