Detecting sensor failures based on environmental and economic parameters of boiler room operation using neural network

Operation of boiler units is often followed by the sensors failure, their readings are not true by any reason. At the Ulyanovsk TPP-1 in January 2021, an experiment was carried out to clear the main technological parameters of the boiler unit No. 1. The statistical data obtained in the experiment formed the basis of the training sample for the neural network. To solve the problem of predicting one of the parameters, it was decided to create a single-layer neural network based on regression of many variables. The content of oxides in flue gases was taken as a predicted parameter. A neural network is a single-layer network with one output neuron and four input neurons. After fully training of the neural network, a prediction accuracy test was performed based on test data. The test prediction error was 0.0076, which indicates the high accuracy of the developed neural network. For the convenience of obtaining predictions using a neural network and outputting additional data (efficiency), a function was developed that takes the following values at the input: natural gas consumption, O$_2$ content in flue gases, steam consumption behind the boiler and temperature of flue gases. Based on the input data, a prediction of the NO$_x$ content in the flue gas is made. This predicted parameter value is compared with the actually measured value, and based on this, it is concluded that the sensor needs to be replaced or calibrated. This function allows improving the existing decision support systems by reducing the percentage of false prompts.