Demand-side flexibility integration into virtual power plant models

Abstract

The present work explores the capacity to estimate and offer demand-side flexibility through centralized management agents such as virtual power plants and demand aggregators, considering a bottom-up approach to characterise flexible loads and the evaluation of the available flexibility individually and under an aggregate approach. In addition, the influence of external factors on the available flexibility is evaluated, considering price scenarios, weather scenarios, market requirements, appliance operating and user comfort constraints when flexibility is estimated. To reach the main objective, four specific objectives were defined, related to: the development of models to characterize flexible loads and their flexibility; the development of a demand aggregation model for flexibility estimation; the development of forecasting models for temperature and solar irradiation; and finally, the economic evaluation of flexibility and its response capacity to flexibility prices under a virtual power plant approach, considering meteorological and stochastic price scenarios. This thesis is composed of 5 chapters, of which three chapters correspond to published or submitted scientific articles, which comprise this thesis.

Methodologically, this work is divided into three stages, following the four specific objectives. The first stage (Chapter 2) focuses on the modelling of flexible loads and demand aggregation models. By exploring residential loads, their properties, and operating constraints, it is possible to establish operative models for each flexible load and subsequently extend them to evaluate the flexibility of each load.

Subsequently, a demand aggregation model was developed for the joint evaluation of flexible loads and their capacity to offer flexibility over time, artificially extending the temporal flexibility capacity of the loads compared to the flexibility of each load individually.

The second stage (Chapter 3) focuses on the development of forecasting models for meteorological conditions, focusing on solar irradiation and ambient temperature. Considering cloudiness as an element correlated with both evaluated elements, a neural network model (LSTM, long short-term memory) was developed. Using this approach, a methodology to determine the probabilities of different cloudiness conditions was developed using Markov models, representing the probability of occurrence of each scenario. In the third stage (Chapter 4), the flexibility models presented in Chapter 2 and the stochastic scenarios introduced in Chapter 3 were added to a virtual power plant model. The purpose was to evaluate the flexibility's responsiveness to different external variables and the flexibility prices. The flexibility model was extended to include new flexible loads and the evaluation of the rebound effect, as a counterweight to the benefits obtained by providing flexibility to the system, making the offered flexibility sensitive to the costs produced by time-shifted energy requirements.

The results of this work permit to use the of flexibility on the demand side when scheduling flexible loads and purchasing energy in day-ahead markets. The Flexibility is evaluated with an adaptive framework, which allows an efficient calculation of the available flexibility and evaluation of different flexibility offers.

It requires a single calculation of both aggregate demand and flexibility over the entire evaluation period, rather than an individual estimation for each time interval. The influence of weather conditions on flexibility as a consequence of thermal energy losses requires an in-depth analysis of temperature conditions.

Consequently, a scenario generation model was formulated based on cloudiness index, using an LSTM model to predict solar irradiance and temperature, characterizing the San Diego climate and generating correlated scenarios. These scenarios were evaluated under a simple virtual power plant model, reducing operational costs compared to a naive forecast, decreasing the number of days with high costs due to unexpected changes in weather conditions.

The integration of flexibility estimation, stochastic weather scenarios, and the flexibility rebound effect produces a model sensitive to flexibility prices, offering a flexibility response similar to a sum of four cumulative normal distribution functions (normCDF).

Two means from the normCDF were related to a sudden change in flexibility: the first mean is linked to the action of flexible loads with minimum and low flexibility costs, while the second mean corresponds to the penalty for discharging electric batteries. The duration of the flexibility rebound time directly affects the available flexibility, increasing flexibility by 30% when doubling the rebound time, while halving it decreases flexibility by 25%.