Method for Generating Labeled Sample Sets for Power System Load Flow Relationship Learning

In constructing power system load flow mapping relationships using machine learning algorithms, the fundamental prerequisite for ensuring the computational accuracy and generalization performance of the mapping model is the availability of a suitably sized, well-distributed, and high-quality labeled sample set that can be supplied in a precursory and efficient manner. Here, the quality of a load flow sample set is defined concretely by three concurrent properties, as follows: physical consistency (satisfaction of Kirchhoff’s and Ohm’s laws), representative coverage of the operational state space, and low inter-sample redundancy. Currently, both online and offline techniques for power flow samples are incapable of efficiently providing large-scale, high-quality power flow sample sets in this sense. To address this, the paper proposes a method for generating power flow sample sets that integrates a physical model of the power grid. This method encompasses the following: non-iterative, high-speed generation techniques for massive load flow samples; partitioned generation and multi-region splicing techniques for large power grid load flow samples; and the design of capacity requirements and quality technical indicators for load flow sample set production. Analytical results demonstrate that the proposed method can efficiently produce high-quality power flow sample sets of appropriate capacity based on actual needs. Case studies on the IEEE 9-bus and 39-bus systems show that sample generation is about 19 and 32 times faster than the whole-network Newton–Raphson method, respectively, for 100,000 samples, and the voltage-band capacity design requires only about 12.6% of the samples needed by uniform sampling for equal boundary-condition coverage, at a comparable learning error.