Machine Learning Approach to Day-ahead Scheduling for Multiperiod Energy Markets under Renewable Energy Generation Uncertainty

In this paper, we propose a day-ahead scheduling method under uncertain renewable energy generation based on a machine learning approach. An aggregator, which has renewable energy generation devices, needs to schedule the energy production and consumption (prosumption) in a situation where the renewable power generation amount is not exactly predicted at day-ahead scheduling. If imbalance, defined as the difference between a day-ahead schedule and a prosumption profile on the next day in the day-ahead energy market, occurs, the aggregator must pay imbalance adjustment costs. As a scheduling method to avoid paying imbalance adjustment costs, we propose a scheduling model by machine learning based on the results of past transactions. We first formulate a problem of constructing a scheduling model as a problem of finding parameters involved in the scheduling model. Next, by introducing a kernel method, we show that the problem of finding the parameter maximizing the mean of profits of past transactions is a concave program. Furthermore, by introducing piecewise affine cost functions, we also show that the problem of finding the parameter can be formulated as a quadratic program. Finally, we show the efficiency of the proposed method through a numerical example.