Prescriptive Trees for Value-oriented Forecasting and Optimization: Applications on Storage Scheduling and Market Clearing

Decision-making in the presence of contextual information is a ubiquitous problem in modern power systems. The typical data-decisions pipeline comprises several forecasting and optimization components deployed in sequence. However, the loss function employed during learning is only a proxy for task-specific costs (e.g. scheduling, trading). This work describes a data-driven alternative to improve prescriptive performance in conditional stochastic optimization problems based on nonparametric machine learning. Specifically, we propose prescriptive trees that minimize task-specific costs during learning, embedded with a scenario reduction procedure to reduce computations, and then derive a weighted Sample Average Approximation of the original problem. We present experimental results for two problems: storage scheduling for price arbitrage and networkconstrained stochastic market clearing, respectively associated with electricity price and load forecasting. The empirical results show significant improvements on deterministic and stochastic lookahead policies, with the relative difference being more pronounced when training data are limited.