Research On Reservoir Flood Forecast Based On Deep Learning

For a long time,the flood disaster in our country has always been one of the most frequent natural disasters that seriously threaten the safety of human life and property.In order to alleviate the negative impact of such problems,China has built a large number of reservoirs with flood control functions.Reservoir flood forecasting,as an indispensable non-engineering measure,plays a key guiding role in flood control decision-making.Reservoir floods are caused by many factors and complex relationships,and it is usually difficult to describe the hydrological process state of the basin completely and accurately.This difficulty has become a constraint to the development of traditional reservoir flood forecasting models.As traditional hydrological forecasting methods have gradually showed many shortcomings such as inconvenient,inefficient,and inaccurate,it is necessary to conduct more in-depth research and more innovative practices on flood forecasting in reservoirs to help flood forecasting in guiding flood prevention and mitigation work play a more positive role.At this stage,the development of artificial intelligence is growing.The reservoir flood forecasting model based on deep learning is a black-box model about the data input-output relationship.The early inflow and rainfall of the reservoir are used as the hydrological driving factors of the model,and the current inflow of the reservoir is used as the corresponding predictive factor.In the study of reservoir flood forecasting based on deep learning,firstly,the gray correlation analysis method is used to calculate the gray correlation degree between the hydrological driving factors and predictive factors of the model,and the hydrological driving factors are initially selected based on the above results,and the selected results are reasonable combination.Secondly,in order to avoid the overfitting of the model due to the noise in the data,the Gamma Test method is used to estimate the noise of different hydrological driving factor combinations,and the driving factor combination with the smallest noise is selected as the final input of the model.For the problem of high computational cost when searching for each neighbor point and solving the corresponding neighbor value in Gamma Test,the KD tree program written based on Python language is used for auxiliary calculation.Thirdly,the PyTorch deep learning framework is used to construct the LSTM-based reservoir flood forecasting model,the CNN-based reservoir flood forecasting model,and the CNN-LSTM-based reservoir flood forecasting model,and the hyperparameters in the trial-and-error method are used for optimization.Different models use different optimizers to optimize the gradient of network parameters,and properly introduce regularization,batch standardization and other methods to improve the models.Finally,evaluate the reservoir flood forecasting capabilities of the LSTM model,CNN model and CNN-LSTM model,including model accuracy assessment and correlation analysis between predicted and measured flows.Based on the above results,the three models are compared and evaluated.This paper takes Wohushan Reservoir as the background,and after the selection of hydrological driving factors,the reservoir flood forecasting models based on LSTM,CNN and CNN-LSTM are established respectively.The concept of deep learning is introduced into reservoir flood forecasting,and the strong adaptive learning ability and high degree of nonlinearity of deep neural networks are used to improve the overall performance of the reservoir flood forecasting model.The results show that the LSTM model is good at describing the overall trend of the flood process,and its Nash efficiency coefficient and certainty coefficient are higher than the calculation results of the CNN model.The predicted flow and the measured flow also show a significant and strong correlation.The CNN model has a more prominent performance in the simulation of flood peak flow,and is better at correctly handling extreme situations when the flood volume changes rapidly.The CNN-LSTM model has the advantages of the previous two models,but its overall prediction accuracy needs to be improved.These three models have their own limitations,and the small sample data also has a negative impact on the effect of the model.However,compared with traditional models,models based on deep learning are more efficient and convenient,and can show a certain degree of applicability and reliability in flood forecasting of reservoirs,and have strong application potential.