Research On Health Condition Assessment Method Of Power Station Equipment Based On Deep Learning

In the new power system,renewable energy units are taking on more and more important tasks of power generation,but the randomness and uncertainty of natural resources are still serious challenges that must be overcome.Coal-fired units need to complete flexibility adjustment to meet the need of depth peak and frequency regulation,but frequent load changes will increase the failure rate.Therefore,no matter for renewable energy units or traditional coal-fired units,it is very important to construct the health state monitor and evaluate system of the key equipment to ensure the safety,efficienicy,cleanness and reliablity of power generation process.In this paper,deep learning theories are used to establish the state prediction model that can fully excavate the time characteristics and internal correlation of operation data in power units.Considering the influence of load changing,a dynamic ensemble prediction model based on deep learning algorithms is proposed.In order to improve the accuracy and timeliness of the fault warning strategy of power plant equipment,a reasonable and reliable fault warning method combining the static threshold and dynamic adaptive threshold of the condition monitoring index is proposed.This method can capture slightly fault characteristics even when operation state changes,so as to guarantee the long-term stable operation of units.The main research work of this paper includes:(1)A dynamic ensemble state prediction model based on the convolutional neural network(CNN),long-short term meomory(LSTM)network,attention mechanism and genetic algorithm(GA)is proposed.CNN completes operation status classification based on characteristic parameters that can characterize the operation status of power units equipment,and outputs soft classification labels as initial weights for model ensemble.LSTM network realizes accurate state prediction by extracting the time information contained in input series.The introduction of AM can strengthen the influence of correlation information.The item of weight biases is set to adjust initial weights,and GA is used to optimize values of biases.(2)A state monitoring index describing the operation state of equipment is established and a fault warning method combining the dynamic threshold and static threshold is proposed.Firstly,the deviation function is constructed based on the prediction error between the ensemble predicted value and the actual observed value.Then the sliding window technique is used to devide the deviation sequence.According to distribution characteristics of deviation sequences in each time window,the state monitoring index and dynamic warning threshold are constructed by introducing the decile and percentile.Finally,the static threshold is calculated based on the prediction deviation of normal dataset,and the combined threshold is obtained by combining the dynamic threshold and static threshold.Therefore,the accuracy and timeliness of early fault warning can be improved.(3)A forced draft fan of the coal-fired power unit and a wind turbine blade are studied as examples.After selecting the feature variable,modeling variables and preprocessing the sample data,the CNN-LSTM-AM dynamic ensemble prediction model and a single LSTM-AM model are established and compared.Results show that the dynamic ensemble model proposed in this paper can mine the information contained in the input sequence more completely and have the better prediction accuracy.The fault warning method based on prediction deviation and quantile is tested based on the fault dataset.Simulation results show that the proposed warning method can quickly and accurately identify the normal and fault state,and timely sending the warning signal even when load changes.Besides,the method is not limited by the application filed,so it is important for the safe and reliable operation of critical equipment.