Research On The Operation Safety Guarantee Method Of Hydropower Units Based On Machine Learning

As the core device in the process of hydropower generation,the stability of the unit not only affects the economic benefits of the power station,but also affects the safety of the power station and staff.Due to the mega-size and complexity of random hydropower units,the safety risks in operation increase accordingly.At the same time,due to the lack of understanding of the operation rules of units and the lack of safety precautions,the stability and safety problems of units have occurred in many power stations at home and abroad.In view of the shortage of prevention and control measures for the current unit safety problems,it is urgent to carry out the research on the method of unit safety operation to avoid the occurrence of emergencies,establish reasonable maintenance intervals and accurately and quickly identify the fault types of units.Here,the unit safety guarantee method is divided into three parts:state evaluation,fault warning and fault diagnosis.Among them,the state evaluation is to describe the unit's health condition and the degree of deterioration quantitatively.The purpose of fault warning is to accurately and quickly identify the abnormal state of the unit and send out an alarm to prevent the expansion of the fault.Fault diagnosis is to identify the current fault type of the unit through the fault characteristics shown by the stability parameters and design the maintenance scheme according to the diagnosis results when determining the abnormal unit.It is difficult to study the stability law from the perspective of mechanism because of the complex structure and various disturbance factors.The monitoring data of units contain a wealth of status information,so it is an effective method to carry out the research on unit security based on the data.Machine learning is a data-driven training model to realize the classification,regression and clustering of data.Its performance is superior and it is widely used.At the same time,the core research questions of state assessment,fault warning and fault diagnosis are consistent with the functions realized by machine learning.According to this,this paper focuses on the problems of state evaluation,fault warning and fault diagnosis,and explores the methods of achieving unit operation safety guarantee based on machine learning and deep learning theory.Aiming at the problems to be solved in the unit safety guarantee,a corresponding model is proposed and the model is tested on the actual unit monitoring system data set or simulation data set.The main contents and conclusions of this article are as follows:(1)Based on a comprehensive analysis of the storage structure and storage strategy of the monitoring database,the concept of a data unit is proposed,which is used as the minimum input unit for various types of security applications to ensure that the data unit contains all the characteristic components of the data.Build metadata on the basis of stored data,treat it as a secondary description of the data and use it as a data unit label to quickly build a data set,and prepare for training models that implement different functions.(2)On the basis of the unit long-term monitoring data analysis and build a variety of test statistics and machine learning model,found that the stability parameter change trend influenced by active power and the working head,but due to the randomness of the signal itself with the inefficiency of condition parameters prediction method,only the combination of historical data and working parameters on the stability parameter prediction effect is good.Aiming at the problem that it is difficult to predict the parameters of pumped storage units effectively due to the frequent switching of operating conditions,a method based on LSTM stability parameter prediction is proposed,and the pendulum degree of the actual units is taken as the target for prediction.The actual tests prove that this method can effectively predict the variation trend of parameters.(3)Aiming at the problem that the current fault early warning method does not respond in time,an early warning method based on energy operator and K-means clustering is proposed.Calculate the Teager energy operator on the collected data to obtain the vibration energy value,and predict the vibration energy trend information through the neural network-under-complete self-encoder integrated model,and use the predicted value and historical sequence as input to perform a clustering model to determine whether there is an abnormal state.So as to realize the early warning of the sudden failure of the unit.The method is verified by the steady-state operation of the real unit and the data of the inserted fault.From the warning results,it can be seen that the warning method proposed in this paper can effectively judge the vibration energy transition,that is,it can effectively warn of sudden faults.(4)In this paper,multiple inherent multiple classification models and combined multiple classification models are constructed for the multi-classification problem of fault types in fault diagnosis,and their classification effects are compared on the simulation feature data set.It can be seen from the test results that the classification effect of the random forest model is the best.At the same time,in terms of the combination mode of the binary classification model,although the one-to-one binary classification combination mode is higher than the one-to-many combination mode in the computational complexity of the model,it performs better than the one-to-many combination mode in the accuracy and precision of the test samples.(5)Aiming at the problem of sample imbalance caused by the scarcity of unit fault samples,a sample generation model based on generative adversarial network is proposed,which realizes the expansion of the sample set by high-precision forging of small samples.The model is tested on the simulation data set,and the results show that the frequency domain and time domain characteristics of the forged samples are very different from the input samples,which can be applied to solve the problem of sample imbalance.