MANY-TASK LEARNING FOR INDIVIDUALIZED CONSUMER POWER PREDICTIONS

Abstract

Short Term Load Forecasting (STLF) aims at predicting power consumption in the hours, days, or weeks ahead. Accurate STLF is important for plant scheduling, financial planning, system security, short-term maintenance, short term storage usage, and the application of demand response strategies, which aim at rewarding reduced power consumption at peak hours. State of the art work uses single task deep learning (STL) for its ability to model the uncertainties in the individualized load. However, the approach can be improved further by combining data from other smart meters. To advance STLF accuracy, this thesis explores the use of transfer learning. Two new transfer learning models are proposed: A hierarchical clustering with population STLF prediction models (HC-P) and a hierarchical clustering with deep multitask learning (HC-MTL). Each of the two hierarchical algorithms cluster similar smart meters into groups based on smart meters’ data representation. The HC-P approach allows smart meters in each group to learn a shared feature representation. The HC-MTL approach uses hard parameter sharing (HPS) schemes for smart meters in the same group and soft parameter sharing (SPS) for smart meters that are unique and different from all other smart meters. The thesis’s additional contributions include two studies for deeper insights into the limitations and opportunities of using transfer learning for STLF. The first study examines the effect of available data on STLF accuracy. The study shows that more smart meter training data helps in improving STLF accuracy, but up to a certain saturation point beyond which, limited performance gains can be obtained. This insight suggests that transfer learning will only make a difference for STLF of smart meters that do not have sufficient training data. To confirm the benefit of transfer learning, a second study examined the effect of getting more data by way of adding data from similar smart meters. Finally, to evaluate the proposed transfer learning STLF approaches HC-P and HC-MTL, experiments were conducted on a dataset consisting of 4225 residential smart-meters and 484 industrial smart meters. Several models were implemented for comparison, including: the state of art STL model composed of 1D-Convolutional Neural Network (CNN) with Gated Recurrent Unit (CNN-GRU), a population prediction model without grouping, a hierarchical random grouping with population prediction models, Autoregressive Integrated Moving Average (ARIMA), and Seasonal ARIMA (SARIMA). The results showed that the HC-P worked best for residential smart meters’ STLF and HC-MTL worked best for industrial STLF. In fact, compared to prior state of-the-art, HC-P provided an accuracy improvement of 2% RMSE and MAE for residential smart meters. For industrial smart meters, HC-MTL provided an improvement of 2.78% and 4.97% in terms of RMSE and MAE, respectively