CN-115688865-B - Long-period memory network industrial soft measurement method for flue gas of desulfurization process

Abstract

The invention relates to the technical field of industrial soft measurement, in particular to a long-short-period memory network industrial soft measurement method for flue gas of a desulfurization process, which comprises the steps of collecting input and output data in the desulfurization process of a power plant to form a historical sample database; preprocessing acquired sample data, namely establishing a robust LSTM model with Huber loss and l 1 regularization, detecting abnormal values of the data, traversing history data by using a sliding window, carrying out difference on one-bit data after the window and statistics in a reference window to obtain a time sequence S1, calculating the upper edge and the lower edge of an S1 box diagram as normal ranges, if the difference between the current value and the statistics in the reference window is not in the normal ranges, carrying out abnormality, and applying the LSTM model to SO 2 concentration prediction in a desulfurization process of a thermal power plant. Simulation results show that the proposed model outlier and non-Gaussian noise have good anti-interference capability, and the influence of redundant variables is reduced, so that the prediction performance of LSTM is improved, and the method has good economical efficiency and applicability.

Inventors

• GUO JUNMEI
• ZHAO LEI
• SUN KAI

Assignees

• 齐鲁工业大学

Dates

Publication Date

20260512

Application Date

20221102

Claims (2)

1. 1. The long-period memory network industrial soft measurement method for the flue gas of the desulfurization process is characterized by comprising the following steps of: step 1, acquiring input and output data in the desulfurization process of a power plant to form a historical sample database; step 2, preprocessing the collected sample data: Step 3, establishing an LSTM model to detect abnormal values, traversing history data by using a sliding window, and making difference between one-bit data after the window and statistics in a reference window to obtain a new time sequence S1, secondly, calculating the upper edge and the lower edge of an S1 box diagram as normal ranges, and finally, if the difference between the current value and the statistics in the reference window is not in the normal ranges, regarding the difference as abnormal; step 4, applying the LSTM model to SO 2 concentration prediction in the desulfurization process of the thermal power plant; the LSTM model is provided with Huber loss and Regularized robust LSTM model, firstly, in order to raise resistance of LSTM to non-Gaussian noise and abnormal value, adopting Huber loss to replace MSE loss of classical LSTM, secondly, in order to reduce complexity of LSTM model, raising generalization performance of model, adding in loss function Performing gradient update using Adam after completing gradient derivation of the proposed LSTM model; With Huber losses and The loss function in the regularized robust LSTM model is ; LSTM at The external state of the moment is In the following The error term of the moment is defined as , 、 、 、 And their corresponding error terms are defined as: Here, the ; When the output door is In the time-course of which the first and second contact surfaces, The weight gradient of (c) is the sum of the gradients at each instant, at instant t, Can be expressed as the gradient of (2) Here, the Final result The gradient of (2) is In the external state at the previous moment, In order to be able to enter the door, As an internal state at the present moment, In order to output the door, the door is provided with a door opening, As an external state at the present moment, 、 、 、 And 、 、 、 Respectively a weight matrix and vector bias, and s and u are respectively the numbers of neurons of the input layer and the hidden layer; for the vector dot-product, Activating a function for the tanh; for the input of the current moment of time, And The final regression coefficient vector and bias, respectively.
2. 2. The method for industrial soft measurement of long-term memory network for flue gas in desulfurization process according to claim 1, wherein in step 2, for the original time series By means of sliding windows A multi-step time series is constructed for LSTM, , Wherein For the length of the original time series, Represent the first The input data of the moment of time, Represent the first The response output of the moment in time, The prediction step size is represented as such, The memory capacity of the LSTM is determined.

Description

Long-period memory network industrial soft measurement method for flue gas of desulfurization process Technical Field The invention relates to the technical field of data-driven industrial soft measurement, in particular to a long-short-period memory network industrial soft measurement method for flue gas of a desulfurization process. Background In modern industrial processes, on-line real-time monitoring and control of key performance variables that ensure process safety, production efficiency and product quality are required. However, some performance variables, such as mud composition, fluid viscosity, and gas concentration, are difficult to measure directly with hardware sensors. The data-driven soft measurement uses historical data collected by a process automation system to build a mathematical model, provides a quick, low-cost and easy-to-implement alternative for real-time monitoring and control of key variables of an industrial process, and therefore, is widely studied and applied. However, the current soft measurement modeling of the industrial process still has the problems that (1) the industrial process generally shows strong nonlinear characteristics due to complex physicochemical reactions and structures, (2) a plurality of redundant variables often exist in the industrial process, so that the complexity of a data-driven soft measurement model is too high, an overfitting phenomenon can occur, (3) the position distribution of each sensor in the process can cause a large time delay between the measurement results of each variable, and (4) the actual industrial process can be influenced by environmental interference, sensor drift and system faults, and non-Gaussian noise and abnormal values exist in the measurement data. In the desulfurization process of a thermal power plant, the concentration of SO 2 in flue gas discharged by a secondary absorption tower is a key performance index related to pollutant discharge standard and energy consumption. However, the index is often unstable and difficult to measure directly with a hardware sensor. Meanwhile, certain online analyzers are often expensive and time consuming to measure. Therefore, it is necessary to design a soft sensor to provide calibration information, improving the reliability of the system. Disclosure of Invention Aiming at the problems, the invention provides a long-period memory network industrial soft measurement method for flue gas in a desulfurization process. The invention provides a technical scheme that the industrial soft measurement method of long-term and short-term memory network for flue gas in desulfurization process comprises the following steps: step 1, acquiring input and output data in the desulfurization process of a power plant to form a historical sample database; step 2, preprocessing the collected sample data: Step 3, a robust LSTM model with Huber loss and l 1 regularization is used for detecting abnormal values of the data, sliding window traversal history data are used for carrying out difference between one-bit data after window and statistics in a reference window to obtain a new time sequence S1, the upper edge and the lower edge of an S1 box diagram are calculated to be used as normal ranges, and finally, if the difference between the current value and the statistics in the reference window is not in the normal range, the current value is regarded as abnormal; and 4, applying the LSTM model to SO 2 concentration prediction in the desulfurization process of the thermal power plant. The LSTM model is a robust LSTM model with Huber loss and l 1 regularization, and firstly, in order to improve the resistance of LSTM to non-Gaussian noise and abnormal values, the MSE loss of classical LSTM is replaced by the Huber loss. Secondly, in order to reduce the complexity of the LSTM model and improve the generalization performance of the model, l 1 regularization is added into the loss function. In step 2, for the original time seriesWith a sliding window m, a multi-step time series is constructed for LSTM,Where L is the length of the original time series, x t represents the input data at time t, d t represents the response output at time t, p represents the prediction step size, and m determines the memory capacity of LSTM. 1. Algorithm foundation 1. LSTM neural network algorithm LSTM was proposed to overcome the gradient extinction and gradient explosion problems of standard RNNs. The information flow is controlled by designing three gating units. Under the control of three gates, the LSTM can forget the previous useless information and keep and transfer the useful information in the memory unit, and has the capability of predicting time sequence. Because of these advantages, LSTM is well studied in the field of soft sensors. The structure of LSTM in the study is shown in FIG. 1. The specific flow of the information processing is that firstly, the memory unit c t-1 at the previous moment forgets some useless informati