Parallel stochastic dynamic programming for long-term generation operation of cascaded hydropower stations

Abstract

Stochastic Dynamic Programming (SDP) for long-term generation operation of cascaded hydropower stations will bring about the curse of dimensionality, resulting in the rapid increase of computational time and the decrease of computational efficiency. Therefore, alleviate the dimensionality problem and improve the computational efficiency are always difficult issues for long-term generation operation of cascaded hydropower stations. On the basis of the parallelism analysis for SDP, a parallel stochastic dynamic programming (PSDP) based on Fork/Join parallel framework was proposed. In this method, all computational tasks for the returns from all discrete combinations in one stage were taken as parent task, which was decomposed into several subtasks by divide-and-conquer method. After this, the decomposed subtasks were solved in different cores respectively for achieving fine-grain parallel computation. The proposed approach was implemented to long-term generation operation of cascaded hydropower stations located on Lancangjiang River, and 3 different schemes with different discrete number of variables were established for testing the computational efficiency in multi-core environment. The result shows that the computational time, compared with serial computation, decreased respectively about 50% in 2-core environment and 70% in 4-core environment, making full use of multi-core resources. In addition, the larger computational scale can reduce more computational time in multi-core environment. Hence, the proposed approach is effective for operation of large-scale hydropower system, and can provide guidance for other applications.