This paper proposes a method for standardized collection of hydropower plant equipment data, and establishes a bi-directional long and short-term memory network (Bi-LSTM) model applying the attention mechanism. After the standardized collection of hydropower plant operation data and feature processing, an equipment fault diagnosis process is established, and a variety of fault pre-processing schemes are formulated according to the actual situation, such as adjusting parameters, distributing loads, hierarchical response and closed-loop feedback, etc. The Bi-LSTM model is also used in the experiments to verify the accuracy of the data collected. The experiments verified that the Bi-LSTM model surpasses the classical algorithms such as SVM, BP and CNN in fault identification accuracy, and its accuracy can reach 92.14% when the training set has 1000 samples. Moreover, the performance test of the system also shows stable response time, high transmission efficiency, and possesses good real-time and scalability. The proposed research can supply theoretical basis and technical route for constructing intelligent and solid housing power system, and promote the management of hydroelectric power station equipment in the direction of intelligent forecasting and automatic maintenance.
