Research On Prediction Method Of Oilfield Water Injection Flow Based On Machine Learning

Accurate prediction of oilfield water injection flow can improve the efficiency of production department testing and adjustment,reduce resource waste,and improve oilfield recovery rate,which is of great engineering significance.In recent years,with the development of machine learning algorithms,more and more machine learning algorithms have been applied to practical industrial control fields due to their advantages of low cost,flexible and variable structure,and high work efficiency.Due to the close connection between oilfield water injection flow related data,oilfield water injection flow related data can be regarded as time series data.Therefore,machine learning time series prediction algorithms can be used for oilfield water injection flow prediction.This article proposes various methods for predicting oilfield water injection flow based on machine learning algorithms.The main work includes the following aspects:Firstly,in response to the problems of poor data feature extraction ability and low data utilization in classical algorithms such as MLP and RNN,this paper proposes a prediction method for oilfield water injection flow based on CNN-Bi LSTM.This method uses one-dimensional convolution and maximum pooling methods based on the characteristics of oilfield water injection related data to extract features from the data,solving the problem of poor feature extraction ability of classical algorithms;Based on the close connection between oilfield water injection related data,reverse calculation has been added to the classic algorithm LSTM,which solves the problem of low utilization of LSTM data and inability to explore the temporal features of data from back to front.Experiments were conducted on two datasets with vastly different styles both domestically and internationally,and the results showed that the proposed method outperforms classical algorithms and is feasible in the field of oilfield water injection flow prediction.Secondly,the CNN-Bi LSTM algorithm is unable to highlight important information and its performance near the inflection point of the water injection flow curve still needs improvement.This article proposes a method for predicting oil field water injection flow based on dual attention mechanism.This method introduces dual attention mechanisms,namely feature attention mechanism and time step attention mechanism,on the basis of the CNN-Bi LSTM algorithm,with the multi head attention mechanism as its underlying operating logic.The essence of multi head attention mechanism is a collection of multiple self attention mechanisms.Each self attention mechanism represents a subspace.Each subspace learns different behaviors,and then combines different behaviors as knowledge,which can improve the generalization ability of the model and prevent overfitting.The feature attention mechanism and time step attention mechanism assign higher weights to the hidden layer states of important features and key time steps,making them more likely to affect subsequent results.Experiments were conducted on two datasets both domestically and internationally,and the results showed that the proposed method effectively tracked the actual curve near the inflection point of the water injection flow curve,solving the problem of CNN-Bi LSTM not being able to detect the changes in the water injection flow curve in a timely manner and not being able to track the actual curve in a timely manner.Finally,in response to the problem of network degradation in traditional deep neural network algorithms,this paper proposes two algorithms,namely,introducing gated quick connections and residual connections respectively on the basis of traditional deep neural network algorithms.Both connections allow the algorithm to continuously learn residuals to optimize the algorithm.However,compared to gated quick connections,residual connections are not limited by the gating unit,and all information always passes through,Able to continuously learn residuals while retaining information from the previous layer of the network.Experiments were conducted on a large volume of water injection development data set of complex fault block reservoirs in the Gulf of Tonkin,and the CNN Bi LSTM algorithm and the traditional depth neural network algorithm were used as the benchmark algorithm.The results showed that the traditional depth neural network algorithm had network degradation.The introduction of gated quick connection and residual connection solved this problem,and the prediction curve was back on track.When residual connection was introduced,the effect was better.