Research On Distributed Stochastic Dynamic Programming Algorithms For Cascaded Hydropower Stations

With the development of the hydro system in China, a large number of large-scale cascade hydropower stations have been put into use. Hydropower station group posses many characteristics such as power station series, large installed capacity, wide transmission range, and its optimization has many characteristics such as high dimension, nonlinearity, multi stage and multi constraint. Stochastic dynamic programming is one of the most widely used algorithms in hydropower optimal scheduling. With the increasing number of the hydropower stations, the computational time increases dramatically, "curse of dimensionality" problem has become more and more prominent. Traditional optimization methods have many limitations, unable to meet the real world scheduling refined requirements. How to solve the hydropower optimal problem more quickly, or find a more efficient method to solve the problem, is of great significance in the current hydropower operation.With the development of high performance computing and cloud computing technology, distributed computing technology with cluster has been widely adopted. The distributed computing technology uses many machines for parallel computing, which provides a new method to improve the efficiency of the calculation. In this paper, distributed stochastic dynamic programming methods for long term optimal operation of cascade hydropower stations on Lancang River based on the high performance and cloud platform are studied; and comparisons between distributed stochastic dynamic programming algorithms on different platforms are performed. The details are as follows:(1) Referring to peer to peer communication dynamic programming algorithm based on MPI, distributed parallel stochastic dynamic programming on MPI algorithm DPSDPoM is implemented. To solve the problems of redundant memory consumption and communications, the multithread hybrid distributed parallel stochastic dynamic programming on MPI algorithm DPSDPoM-MT is proposed. Experiments show the scalability of DPSDPoM, which can better assign tasks to computing nodes than traditional methods. DPSDPoM-MT algorithm is superior to DPSDPoM in terms of computational efficiency and memory consumption.(2) Distributed parallel stochastic dynamic programming based on Spark algorithm DPSDPoS is proposed, and two implementations for the algorithm are introduced. The algorithm transforms calculation model of stochastic dynamic programming to data processing model which take full use of cluster resources, and has a effective backup and fault tolerance mechanism and reduces, which has significant advantages. The Experiments show that, compared with DPSDPoM, DPSDPoS is more efficient and extensible; but the framework overhead of DPSDPoS is little high, and memory consumption is more serious than DPSDPoM.(3) Algorithm analyses point out that DPSDPoS is more suitable for the cluster which is composed of a large number of normal compute nodes; and analyses for model applicability point out that the two kinds of distributed algorithms have different degree of applicability for the dynamic programming and its dimension reduction model for the long-term optimal operation of cascade hydropower stations.