基于机器学习的黑龙江省强降水致灾预估方法研究

By utilizing heavy rainfall disaster and daily precipitation data from various counties in Heilongjiang Province from 1984—2019，and by using logistic regression and Long Short Term Memory Network（LSTM），disas- ter prediction models for heavy rainfall are established in Heilongjiang Province，including the Da Xing'an Moun- tains，the Xiao Xing'an Mountains，the Songnen Plain，the Sanjiang Plain，and the southeastern half mountainous area. Through machine learning， it is found that the optimal observation days for judging whether heavy rainfall caused disasters in Heilongjiang Province and 5 regions are 4~ 6 days，and the optimal daily precipitation threshold is 16~ 20 mm. Comparing the performance of four models which are the fully connected logistic regression model，the partially connected logistic regression model D that prioritized dates，the partially connected logistic regression model S that prioritized sites，and the LSTM model，the first three logistic regression models showed little difference in per- formance，with the fully connected model performing the best in comparison. The accuracy，precision，recall，and F1 scores of the fully connected logistic regression model in most regions are all above 0. 7. The LSTM model only out- performed the logistic regression models in the Da Xing'an Mountains.