Periodic Reporting for period 1 - iFLEX (Intelligent Assistants for Flexibility Management)
Reporting period: 2020-11-01 to 2022-04-30
The iFLEX project goes beyond the-state-of-the-art in short term load forecasting by developing hybrid approaches that combine machine learning with physics-based modelling. Combining the best parts of physics-based modelling and machine learning makes it possible to achieve accurate and robust models, which can be also replicated for new consumers with minimum efforts. Two novel hybrid modelling approaches have been investigated in the WP3. The more mature approach, documented in D3.1 utilizes machine learning for modelling the baseline consumption and physics-based models to modify the ML forecast according to the law of energy conversion. Physics based greybox models with few parameters learned from data are also applied for modelling the indoor temperature during the DR event and the total heating demand of the building. Scientific publication of this approach is planned for the next phase of the project. The second innovative approach studied in the project utilizes a physics-based simulator to train a common Artificial Neural Network (ANN) model with a large amount of data from different buildings. The idea is that this forces the ANN to learning the underlying physics and therefore provide more robust results when compared to typical ML approaches. This innovative approach is published as a scientific journal (Kannari et al., 2021).
Artificial Intelligence for automated decision making in flexibility management
The iFLEX project goes beyond the-state-of-the-art in AI-based consumer flexibility management by developing data-efficient and safe approaches and demonstrating them in operational environment. The hybrid Deep Reinforcement Learning approach developed in the project utilizes the consumer digital twins to plan and optimize control actions. This allows the solutions to be much more efficient with respect to the interaction with the real world (i.e. data-efficiency is similar to supervised learning instead of reinforcement learning). Robustness is provided with three layered control architecture that builds upon existing building and smart home control systems, extending them with new expert rules for verifying the action proposals made by the model-predictive control system. The initial prototype of this approach is presented in D3.7. A journal publication about the initial solution at TRL 5 is published in IEEE Access (Kiljander et al., 2021)