Deep Learning For Load Forecasting With Smart Meter Data: Online And Federated Learning
Abstract
Electricity load forecasting has been attracting increasing attention because of its importance for energy management, infrastructure planning, and budgeting. In recent years, the proliferation of smart meters has created new opportunities for forecasting on the building and even individual household levels. Machine learning (ML) has achieved great successes in this domain; however, conventional ML techniques require data transfer to a centralized location for model training, therefore, increasing network traffic and exposing data to privacy and security risks. Also, traditional approaches employ offline learning, which means that they are only trained once and miss out on the possibility to learn from new data. Online and Federated Learning are among the potential solutions to alleviate the mentioned concerns. Online models learn from data as they arrive while Federated Learning (FL) approaches train a single ML model in a distributed manner without requiring participants to share their data.
Consequently, this thesis investigates Online and FL for load forecasting with smart meter data. Deep learning typically requires large and diverse data streams; however, this data may not be readily available due to data collection issues/expenses, privacy and security concerns, or other reasons. Therefore, Recurrent Generative Adversarial Network is designed for generating realistic energy data.
To enable continuous learning from newly arriving data and adapting to new patterns without the need to re-train the model, a novel Online Adaptive Recurrent Neural Network (RNN) is proposed. RNN is employed to capture time dependencies while the online aspect is achieved by updating the RNN weights according to new data. The results show that the proposed approach achieves higher accuracy than the standalone offline approaches and other online algorithms.
For FL, an approach based on FedAVG was designed first: this synchronous approach waits for all clients to complete training before the server aggregates weights. Next, FedNorm, an asynchronous approach for FL, is proposed: it aggregates updates without waiting for lagging clients. To achieve this, FedNorm measures the clients' contributions considering similarities of local and global models as well as the loss function magnitudes. The experiments demonstrate that FedNorm achieves higher accuracy than seven state-of-the-art FL approaches.