A Reinforcement Learning-Based Intelligent Scheduling Method for Defective Materials in Power System

Abstract

The invention discloses an intelligent scheduling method for defective materials in a power system based on reinforcement learning. The decision-making process solves the material storage dynamic scheduling problem; lists the Bellman equations for grid-defective materials and selects a solution strategy; modifies the Bellman equation into a data-driven online update form, and determines scheduling actions based on the ε greedy strategy. The invention proposes to solve the joint control and scheduling problem of emergency materials in power system based on Markov stochastic process and reinforcement learning, and the "end-to-end" algorithm does not predict demand, but directly makes inventory control and scheduling decisions; It has been verified, with good convergence and gain, which proves the usability and practical value of the method.

Description

technical field [0001] The invention relates to the technical field of power grid and artificial intelligence scheduling, in particular to a method for intelligent scheduling of defective materials in a power system based on reinforcement learning. Background technique [0002] Statistical optimization method: Model the distribution of various urgent needs according to statistical laws, and calculate the statistically average optimal warehouse distribution through centralized mathematical modeling. [0003] Data prediction method: Based on the idea of ​​data analysis and mining in each area, using artificial intelligence and machine learning methods, build a time series model (sequence-to-sequence model) for the different needs of each area, so as to carry out time series analysis. Prediction; then, on the basis of prediction, centralized layout and optimization of warehousing system and scheduling. [0004] For the statistical optimization method, this method needs to have...

AI Technical Summary

Problems solved by technology

[0004] For the statistical optimization method, this method needs to have complete statistics for all demand distributions in the region. At the same time, every time a state transition or emergency event occurs, the optimal distribution needs to be recalculated, which consumes a lot of computing resources and responds slowly. Limitations: For the data prediction method, the traditional feature selection is usually based on the feature ranking method, according to the importance and correlation of each feature calculated, and take the top k features as the input of demand forecasting. The biggest disadvantage of this method is that Selecting k features with the greatest importance and correlation cannot well represent the global information of the system, so it cannot provide the most abundant information for the prediction system; at the same time, because the predicted result is not the final result, according to the predicted result we will Perform secondary calculations to obtain scheduling and control schemes. The multi-step framework will cause errors to accumulate and lead to deviations in the final results

Images