Research On Reservoir Operation Function Based On Long Short-Term Memory Network

At present,the implicit stochastic optimization operation has been widely used because it implicitly considers the uncertainty of future reservoir operation,and there are no difficulties similar to the explicit stochastic optimization operation calculation.In the implementation process of the implicit stochastic optimization operation method,the multi-year optimal operation plan of the reservoir needs to be calculated according to the deterministic optimization model and the corresponding solving algorithm.The multi-year optimal operation process obtained by the above solution is mathematically a series of discrete data to be fitted.There are many ways to fit these data.The simplest and most common is to use linear regression to fit.However,the operation function is actually more complicated in certain periods,and it is also considered to use a nonlinear function to fit.There are many expressions of nonlinear operation functions,and the typical representative of them is the BP neural network.Its powerful nonlinear mapping capability and flexible network structure can arbitrarily fit the input and output relations of the reservoir system at various periods.However,when the BP neural network is used to extract the reservoir operation rules,the assumptions about the samples are independent of each other,and the time correlation of the samples is not taken into account.At present,when processing such sequence data,the use of recurrent neural networks can avoid these shortcomings of feedforward neural networks.Therefore,this paper proposes a research method of reservoir operation function based on long-term and short-term memory neural network(LSTM).As an improved version of traditional RNN,it has more powerful performance.This method also considers the nonlinear correlation and time characteristics of the multi-year optimized operation scheme: in terms of nonlinearity,each gating unit in the network uses a sigmoid nonlinear activation function to enhance the fitting of the sample’s nonlinear relationship Ability;in terms of time series,the introduction of self-circulation weights enables the establishment of cyclic connections between cells,thereby dynamically changing the cumulative time scale,so that the cells have long-term and short-term memory functions.Based on this,this paper uses LSTM to realize the characteristics of any complex nonlinear mapping function,time series learning ability and high prediction accuracy to extract reservoir operation rules,and apply it to Danjiangkou Reservoir instance.Experimental research shows that the fitting method proposed in this paper is feasible and effective.Moreover,by setting a linear operation function and a BP operation function control group,it is found that the operation function fitted by the LSTM neural network can achieve better benefits in the simulation operation.