The Application of Federated Learning in the Prosumer-centric Flexibility Market

Household-level flexibility has shown great potential, such as mitigating the uncertainty of renewable generation, avoiding congestion issues, and reducing local grid operation and planning costs. However, there are two major challenges to activating scattered household flexibility from prosumers. One is the lack of an incentive mechanism for prosumers. Though the flexibility market is one of the fairest solutions to gather and utilize household flexibility, it barely considers prosumers' interests. The existing market designs often treat prosumers as price takers and passive service providers. The other barrier is the rising privacy concerns. The prediction and control of flexibility need advanced data-driven models. And data-driven models require the collection and analysis of historical consumption data. At the household level, the consumption data contains sensitive information, such as users' behaviours, consumption patterns, and household occupation rates. Therefore, this thesis presents a prosumer-centric flexibility where prosumers' privacy is protected via the Federated Learning (FL).

In this thesis, the "prosumer-centric" concept is defined in two aspects. First, the prosumers are no longer price takers. With the Geometric Brownian Motion (GBM) method and Monte Carlo simulation, they can negotiate the price paths with Distribution System Operators (DSO). Second, prosumers decide whether or not and how much flexibility that they want to trade instead of passively accepting orders from aggregators or DSO. Incorporating Home Energy Management Systems (HEMS) with neural network (NN) models, the prosumers trade flexibility without violating their comfort levels.

As for privacy concerns, the FL framework is a powerful tool. It decouples the data acquisition from NN model's training process. The raw training data stays locally with prosumers. To comprehensively analyze the FL model's performance in electricity consumption prediction, we tested four different NN structures in three experiments: accuracy, robustness to data quality, and scalability. Furthermore, we push the traditional FL into a fully decentralized FL to avoid single point failure where there is no central server to collect models' updates. Additionally, the Deep Leakage from Gradients (DLG) attack is analyzed. It can recover the original training data from shared model updates like gradients. As a defence, we utilize Secure Aggregation (SecAgg) to prevent attackers hijacking shared gradients.

At last, we examine the FL models in a scaled prosumer-centric flexibility market with 200 households. The FL is used to solve data isolation to allow participants benefit from others' information without intruding their privacy. The simulation shows that the FL framework has a positive influence on flexibility trading. From the sellers' point, prosumers have more financial gains due to preciser price and demand prediction. As for the buyer, DSO can achieve better load shifting because more flexibility is activated from increasing succeed flexibility trades. Additionally, we observe a phenomenon termed as "competitive disadvantage" in FL collaborations. In the competitive market context, some participants benefit their competitors instead of themselves by joining the FL collaboration. It presents a dynamic relationship of competing and cooperating in our study case.

In conclusion, this thesis explores the application of federated learning in a prosumer-centric flexibility market. It addresses the conflicts between privacy and data-driven models' requirements. The FL framework is tested under various study cases and evaluation metrics. It is proven that the FL can bring a balance between data-driven model's performance and prosumers' privacy. The concept of "prosumer-centric" opens a new gate for flexibility market design. It proves that the communication can be bi-directional or bottom-up. The DSOs can still benefit from household flexibility without treating prosumers as passive service providers. The findings in this thesis are a starting point for a wider discussion and exploration of prosumers. As the smart grid is developing, the role of prosumers is more and more important. It is important to think from the prosumers' perspective and consider their interests and privacy.