Monitoring and Event Detection for Flexibility Management in Cyber-Physical Energy Systems

Growing concerns about climate change and dependence on finite fossil fuels from authoritarian regimes are motivating an accelerated transition towards sustainable energy systems. The development is characterized by increasing integration of distributed renewable generation and electrification of several sectors, which puts stress on electrical distribution grids. A cost- and resource-efficient strategy to relieve grid infrastructure is the flexible operation of distributed energy resources (DERs) such as solar plants, heat pumps, and electric vehicles. Planning, realization and verification of such local flexibility requires closed-loop integration of physical processes with digital systems on DER and distribution grid level. The associated increase of operational complexity, new events and failure causes, and emerging data and information needs set new situational awareness requirements on asset and grid level.
This thesis studies the deployment of machine learning (ML)-based predictive analytics for addressing operational awareness requirements of DERs and distribution grids that emerge from local flexibility and the associated digitalization. In this context, three research topics are addressed based on seven research articles.
The first topic addresses the question “What are the operational challenges of local flexibility and the associated digitalization that set new requirements on the situational awareness for DERs and distribution grids?”. Flexibility realization can be subject to weather- and consumer-induced uncertainty, computational limitations and data manipulation. Thus, forecasts for flexibility scheduling should be computationally lightweight and robust against manipulations, apart from being accurate. While digitalization enables a broad spectrum of flexibility mechanisms, the associated increasing load activity and emerging cyber threats require new real-time monitoring and event identification solutions: Low voltage (LV) state estimation should account for possible flexibility-induced stochasticity, which includes uncertainty quantification and incorporation of price signals and other data sources. To increase awareness of distribution system operators (DSOs) about flexibility activations from mechanisms without their active involvement, flexibility activation identification should be investigated. Finally, the risk of coordinated attacks against fleets of flexible resources requires advanced attack identification concepts for DERs, which includes integrated evaluation of cyber network traffic and physical process data.
Within the second research topic, the question “How do weather- and consumer-induced stochasticity, computational constraints and cyber attacks impair flexibility realization, and how can data and modeling strategies based on predictive analytics enable or facilitate computationally efficient, cost-optimal and cyber-secure usage of DERs as flexible resources?” is addressed. A systematic evaluation of forecasting strategies for flexibility schedule optimization of a residential photovoltaic-battery system reveals that unpredictable weather- and consumer behavior reduces financial rewards for prosumers by 5-10 percentage points compared to a theoretical optimum. Moreover, it is shown that ML-based forecasting enables such near-optimal cost-savings while being computationally lightweight, applicable with almost no data history, and robust against weather-input manipulations. In contrast, the results demonstrate that scheduling flexible assets based on manipulated price signals can turn savings into additional costs and peak shaving- into peak-reinforcing behavior, which showcases potential consequences of cyber-physical attacks. As means for cyber-secure DER operation, cyber-physical event identification concepts are investigated. It is shown that the joint evaluation of cyber network and physical process data applying supervised ML improves the event identification performance and enables attack and fault detection in one holistic approach. To leverage these benefits while meeting practical requirements such as independence of scare historical event observations and human-verifiable predictions, the data-driven Cyber-Physical Event Reasoning System CyPhERS is developed. It is demonstrated that CyPhERS can generate informative and human-interpretable event signatures which can be evaluated to infer event information such as occurrence, type, affected devices, attacker location, and physical impact for both known and unknown attacks and faults.
The third research topic is concerned with the question “How does local flexibility affect monitoring of distribution grids, and how can data and modeling strategies applying predictive analytics facilitate effective situational awareness for DSOs under high shares of flexible resources?”. On the basis of a systematic evaluation of several flexibility scenarios it is demonstrated that local flexibility can introduce aleatoric uncertainty to data-driven LV state estimation. By applying a Bayesian neural network and adding real-time power and voltage readings from secondary substations to the inputs, flexibility-induced uncertainty can be reliably quantified and partly counteracted. To further support the operational awareness of DSOs, a data-driven flexibility activation identification concept is proposed. It is demonstrated that flexibility activation identification in aggregated load data is feasible while accounting for practical requirements such as real time detection and handling of multiple and partly unknown load-altering event types by applying unsupervised anomaly detection and open-set classification.
The results of this thesis demonstrate how ML-based predictive analytics can be leveraged to provide solutions to operational challenges for DERs and distribution grids in the context of local flexibility and the associated digitalization, while respecting practical requirements such as computational efficiency, prediction verifiability and use of publicly available tools.