Aggregated load forecast and control for creating alternative power system resources using thermostatically controlled loads

Abstract

Power systems are evolving and are trying to use loads and communication infrastructure as a way to compensate the system generation for peak load shaving and ancillary services. Thermostatically controlled loads (TCLs) have electric operational flexibility that makes them suitable for the compensation via direct load control. Aggregated load forecast and control are critical to transform "TCLs with uncertain demand" into "manageable power system resources." The forecast provides predictive information about baseline loads showing normal energy demand of consumers. The control makes TCL aggregations fully responsive to its upper-level requests, which are the focus of this dissertation. This dissertation investigates two multi-horizon forecast engines for aggregated TCLs: (I) wavelet decomposition-based neural networks with K-means clustering (WNN), and (II) a convolutional neural network-based method (CNN). WNN demonstrates a way to achieve accurate aggregated TCL forecasting by using traditional NNs with proper structure design and predictor selection. CNN provides TCL aggregations a generic forecaster without the need to extract specific predictors. The two forecast engines were studied with different types of TCLs and different aggregation sizes, resulting in an enhanced performance when compared with benchmark algorithms. The dissertation also theoretically derives and analyzes aggregation effects showing the dependence between forecast accuracy and aggregation sizes. Additionally, a scalable forecast mechanism is proposed to aggregate forecasts from individual forecast engines, enabling fast forecast deployments with improved accuracy. The proposed mechanism relies on a bottom-up approach with a new Markov-based error reduction algorithm, leading to 20-80% performance improvement when the method is tested with different forecast horizons. Aggregated TCLs can be controlled to join electricity markets with a central hub, virtual power plant. However, there are two main concerns for aggregated load control. The first is associated with the dispatchability of virtual power plant against normal energy demand of individual TCLs in uncertain time-variant environments. The second refers to the tremendous increase of communication and computational requirements needed to perform direct load control on a large population of TCLs. This dissertation introduces aggregators between the direct load controller and TCLs with a novel robust control mechanism to reconcile these concerns. The control is implemented with two layers: the upper layer suppresses the control payback effect with a quadratic optimization model, and the lower layer addresses the power trajectory tracking with a novel payback tracking error model. The control method needs minimum sensing infrastructure since it requires power data only at the aggregation level. Simulations showed a robust reference power tracking characteristic of the aggregator with a percentage root mean squared error between 3.33% - 5.69% under uncertain time-variant environments. The continuous responsiveness indicates that the aggregators manage to convert the aggregated TCLs into "manageable resources" that ensure the dispatchability of virtual power plant. This approach could allow modern power systems to implement new trends such as "load following generation" which could save resources and reduce the environmental impact of power generation and distribution.