WP4 explores efficient solutions and concepts for sharing data for RES forecasting and derived use cases. We interviewed, DTU, the WP leader, to dive into these concepts and understand Smart4RES approach.
Q1. In a decentralized framework where agents share local information generated from diverse sources such as RES power plants, meteorological stations, etc., which method do you propose to optimize collaborative forecasting while preserving privacy?
Smart4RES aims at proposing and benchmarking novel approaches for collaborative learning and forecasting. In such a decentralized framework, local agents (RES producers, met stations, etc.) *do not* share their local data directly, but instead perform local computations based on their data, to help others to improve their forecasts. Various approaches are considered in practice, inspired by the latest advances in distributed optimization, machine learning and signal processing. Specific methods involve the well-known ADMM approach (Alternating Direction Method of Multipliers) and Mirror Descent approaches. The originality of our approaches is that they are developed for online learning (i.e., computing on the fly and adapting to change of conditions), possibly with asynchronous updates, and finally to accommodate privacy preferences of the agents involved to make sure their original data cannot be inferred. Various architectures are also considered, from centralized to fully peer-to-peer, and with potential fusion centers as buffers. Empirical investigations for wind and solar power forecasting so far show great promises, with increased forecast accuracy for nearly all agents involved.
Q2. Can you explain why your concept of data markets for RES forecasting is an essential tool for stakeholders who dispose of valuable data streams?
As information and communication technologies continue to improve, collecting and extracting value from data will become cheaper and easier. Data sharing between different owners has a high potential to improve RES forecasting skill at different time horizons (e.g., hours-ahead, day-ahead) and consequently the revenue from electricity market players. However, economic incentives, trough data monetization, are fundamental to implement collaborative forecasting schemes since RES agents can be competitors, and therefore unwilling to share their confidential data without benefits. A side benefits is that data monetization will trigger novel business models incentivising to deploy new measurement and forecasting technology. This data market should operate in a way that, after some iterations, agents realize which data is relevant to improve its gain, so that sellers are paid according to their data. The preliminary results in Smart4RES show that data markets can be a solution to foster data exchange between RES agents and contribute to reduce imbalance costs in the electricity market. Eventually we aim at merging the two concepts of collaborative learning and data markets so as to incentivise and reward shared information and computation alike.