| CPC G06N 3/084 (2013.01) [G06N 3/048 (2023.01)] | 5 Claims |
|
1. A method for predicting a flue dust concentration, comprising:
cleaning raw data, wherein the raw data comprises coal quality data and quantity data of each batch of coal fed into a furnace, as well as hourly flue dust emission amount monitored by a continuous emission monitoring system (CEMS);
calculating a flue dust emission amount of each batch of the coal fed into the furnace based on hourly coal consumption of a unit, and normalizing the coal quality data, the quantity data, and the flue dust emission amount of each batch of the coal fed into the furnace to obtain a data set;
constructing a prediction model, training the prediction model by using the data set to obtain a trained prediction model, and predicting the flue dust concentration by the trained prediction model; and
comparing the hourly flue dust emission amount with the flue dust concentration predicted by the trained prediction model to determine whether the CEMS is abnormal; and in response to the CEMS being determined to be abnormal, inspecting and repairing, by maintenance personnel, the CEMS;
wherein the prediction model is a backpropagation neural network (BPNN), the BPNN comprises an input layer, at least one hidden layer, and an output layer, a weight initialization mode of the BPNN is Glorot uniform distribution initialization, a threshold is initialized to 0, and a rectified linear unit (ReLu) function is selected as an activation function for the BPNN;
wherein the training the prediction model by using the data set specifically comprises:
dividing the data set into a training data set and a test data set according to a set ratio;
inputting the training data set into the BPNN, and calculating inputs and outputs of the at least one hidden layer and the output layer through forward propagation;
calculating an error of a predicted result of the output layer by using a mean square error (MSE) formula, wherein the MSE formula is expressed as:
 where mse represents a mean square error, y represents a real value, ŷ represents a predicted value, and n represents a total sample number;
performing backward propagation on the error, optimizing model parameters based on an adaptive moment estimation (Adam) optimizer through formulas below, adjusting a connection weight and a threshold between the at least one hidden layer and the output layer:
 where mt-1 and vt-1 represent first-order moment estimates before iteration, mt and vt represent first-order moment estimates after the iteration, β1 and β2 represent exponential weighted average parameters, dk represents a gradient of a weight w and a threshold b, a represents a learning rate, E represents a constant to prevent a denominator from being 0, wt-1 and bt-1 represent a weight to be trained and a threshold to be trained respectively, wt and bt represent a trained weight and a trained threshold respectively;
repeating the steps from the forward propagation to the backward propagation until a preset number of training iterations is reached; and
evaluating a trained BPNN as the trained prediction model by using the test data set;
wherein the evaluating a trained BPNN by using the test data set specifically comprises:
after training the prediction model, inputting the test data set, obtaining a predicted value by forward propagation calculation, and evaluating a predicted result by a MSE, a root mean squared error (RMSE), a mean absolute error (MAE) and a determination coefficient R2, with calculation formulas as follows:
 where SSR represents a regression sum of squares, and SST represents a total sum of squares.
|