UC Santa Cruz

The hidden information and characteristics embedded in electrical load profiles are indispensable for the effective planning and operation of power grids. Load forecasting plays a vitally important role in many applications for the electric industry, e.g., energy generation and transactions, load shedding and restoration, as well as infrastructure expansions. Based on historical data, accurate load forecasts provide a good reference for the needed load demand, which can increase the efficiency and revenues of the electricity generation and distribution companies. In parallel with load forecasting, load monitoring can identify various types and statuses of loads by disaggregating the total power consumption into individual appliance levels. A scientific procedure of load monitoring facilitates the establishment of user-profiles, power usage habits, and peak load shifting. This is beneficial for both the end-users and utilities, improving the overall efficiency of the power network.

This research encapsulates three pivotal works focusing on enhancing the accuracy and efficiency of predicting and monitoring electrical loads. The first work introduces an attention-based neural load forecasting model that utilizes an encoder-decoder RNN and BiLSTM for dynamic feature selection and similar temporal information adaptively, showing superior performance in short-term load forecasting tasks. The second piece builds upon this, offering a unifying framework that integrates time-varying feature weighting, hierarchical temporal attention, and feature-reinforced error correction. This framework has shown outstanding efficacy in electric load forecasting on public datasets. The third work pivots to Non-Intrusive Load Monitoring (NILM). It introduces MATNilm, a multi-appliance-task framework, which efficiently disaggregates total power usage with limited labeled data. It employs a training-efficient sample augmentation (SA) scheme and a shared-hierarchical-split structure with a two-dimensional attention mechanism, enhancing the recognition of spatio-temporal correlations among all appliances. Collectively, these contributions underscore the integration of advanced deep learning techniques to improve the efficiency, reliability, and accuracy of load forecasting and monitoring in power systems.

Extensive numerical simulations reveal that our proposed load forecasting framework surpasses several existing forecasting methods. We emphasized the critical roles of both the feature weighting mechanism and the error correction module in securing this superior performance. In relation to the NILM task, even with just a single day's training data and limited appliance operation profiles, our SA algorithm demonstrates test performance comparable to models trained with a comprehensive dataset. In tandem with our proposed model structure, the simulation results highlight that our approach offers a notably enhanced performance against numerous baseline models, reducing relative errors by over 50% on average.