基于并行动态规划的水库群优化

Parallel dynamic programming for multi-reservoir system optimization

为探索高性能并行计算在水库群优化问题上的应用,该文以经典四水库问题为例,构建多维动态规划模型,采用主从模式策略对动态规划程序进行并行化,利用高性能并行计算机(多达300个核)对该问题求解,得到了不同核数的计算时间、加速比以及并行效率。结果表明:借助分布式计算,动态规划求解水库群优化的计算时间能够有效缩短,加速比将随核数增加进一步提升,并行效率减少趋势缓慢。未来工作需借助分布式内存来克服动态规划的内存过大问题。

The paper shows how high performance computing can solve the multi-reservoir system optimization problem. A multi-dimensional dynamic programming （DP） model is developed for the four-reservoir problem. Then, the master-slave parallelization strategy is used to parallelize the serial DP algorithm. A high performance parallel computer is then used to solve the four-reservoir system with various numbers of cores （up to 300 cores）. The results show that distributed computing effectively shortens the computation time for the algorithm, the speedup further increase as the number of cores increases and parallel efficiency decreases very slowly. Future work should use the distributed computer memory to alleviate DP＇s large computer memory requirements.