Highly accurate oil production forecasting under adjustable policy by a physical approximation network

Abstract

Petroleum is widely used as one of the important energy sources. Engineers need to continuously adjust strategies to maximize oil recovery during development. Then strategy adjustment requires a lot of cost, so it is necessary to select the optimal strategy through production forecasting. In this paper, a self-defined Double-channel Heterogeneous Dynamical Graph network (DHDG) is proposed deal with this problem. It can receive different development strategies predict future production indices base on that. Particularly, the concept of graph network is used for the first time to simulate the well pattern and predict production of all wells in the block at same time. At the same time, in order to fully fit the seepage process between wells, we complicate the edge features of the graph, so that it has enough capacity to express the state information of the formation between wells. Finally, we amplify the features of network output through engineering calculations, which further ensured the integrity of the information transmitted in the recursive process. The example results in two benchmark reservoirs show that it is accurate enough to predict oil production rate in future 600 days and more accurate than LSTM networks.

Keywords

• Production prediction
• Machine learning
• Adjustable policy