Privacy-Preserving Probabilistic Forecasting of Short-Term Net Load: Exploring the Accuracy– Privacy Trade-off

Ehsan  Razavi

An ongoing increase in renewable energy systems deployment and a push for efficient smart grids promoting energy sharing are reshaping energy management systems. This transformation faces challenges in low-aggregate energy trading, with load forecasting being a prime example. Furthermore, as energy trading becomes more prevalent, securing the sensitive data of participants becomes a priority, making privacy preservation paramount. This thesis, firstly, emphasises on a shift from load to net load forecasting, incorporating consumption and photovoltaic generation forecasts. In this respect, a long short-term memory (LSTM) model is utilised to enhance the accuracy of net load forecasting by considering diverse household load profiles and spatial dependencies. Secondly, the research contributes to the probabilistic load (and net load) forecasting. Unlike traditional point load forecasting, probabilistic load forecasting addresses complexities introduced by load and rooftop generation behind-the-meter. A Bayesian neural network (BNN) is employed to quantify uncertainties in the forecasting approach. With these contributions in place, the thesis thereafter introduces its central contribution: a novel privacy-preserving model for short-term (intraday level) probabilistic load forecasting involving a specialized differential privacy model. Designed to ensure data confidentiality, it incorporates the Laplace mechanism, injecting noise into forecast means and perturbing standard deviations to maintain a 95% confidence interval. This model employs a utility approach to balance data accuracy and individual privacy, ensuring it remains useful for users such as grid operators and retail providers. Additionally, the research investigates privacy constraints, introducing concepts of full and limited consent levels among customers. Recognizing varying levels of willingness for data-sharing within communities, this work explores opportunities to achieve the best tradeoff considering individual preferences around the privacy concept. The outcomes of this thesis, supported by a thorough set of simulations, reveal the impact of behind-the-meter rooftop generation on net load forecasting errors. Notably, an observation within a dataset in Australia indicates that the highest forecasting error is likely to occur during morning or evening hours, corresponding with the periods when PV generation is increasing or decreasing at its highest rate. The simulation results also indicate that using data with finer granularity can help reduce forecasting errors. Moreover, this thesis presents findings that satisfy specific privacy and accuracy requirements in probabilistic forecasting. Furthermore, by considering the community’s openness to disclose its load profile, the pathway to an optimal trade-off between privacy and accuracy in net load forecasting, based on data analysis, is demonstrated.

Privacy-Preserving Probabilistic Forecasting of Short-Term Net Load: Exploring the Accuracy– Privacy Trade-off

Files and links (1)

Abstract

Details

Metrics