Temperature monitoring and regulating method for smoke washing and heat eliminating device based on LSTM (Long Short Term Memory)

Abstract

The invention belongs to the technical field of environmental protection, and particularly relates to a method for monitoring, regulating and controlling the temperature of a smoke washing and heat eliminating device based on LSTM (Long Short Term Memory), and the method comprises the steps: obtaining original historical data through sensing equipment disposed on the smoke washing and heat eliminating device, and obtaining the temperature of the smoke washing and heat eliminating device through kernel function mapping and an attention mechanism; the most relevant strengthening characteristics of the water consumption and the temperature change are effectively excavated, and the data value is improved. The LSTM model established based on the temperature state time dependency relationship can accurately capture the time sequence change of the features, and the output local time sequence features are more targeted. And a prediction model of the water consumption required by cooling is constructed by weighted fusion of the enhanced features and the local time sequence features, so that the accuracy and reliability of temperature monitoring of the smoke washing and heat eliminating device are improved, and the method has important significance in the technical field of environmental protection.

Description

Technical Field

[0001] The present invention belongs to the field of environmental protection technologies, and particularly relates to a method for temperature monitoring and regulation of a flue gas washing and heat dissipation device based on LSTM. Background Art

[0002] During industrial production processes, especially in some high-temperature production links in industries such as metallurgy, chemical engineering, and energy, the temperature of exhaust gases is usually relatively high. If these exhaust gases are directly discharged into the atmosphere without purification, cooling, and other treatments, they may cause great pollution to the environment. Therefore, it is necessary to cool and purify them. Existing flue gas washing and heat dissipation devices usually adjust the temperature based on simple valve control or a fixed spray volume, but these methods cannot make dynamic adjustments according to the real-time changing flue gas temperature and flow rate. Therefore, there are often phenomena of excessive or insufficient water consumption, which not only increases the waste of water resources but also makes the energy efficiency and environmental protection effect of the equipment unable to reach the optimal state. During the process of high-temperature flue gas treatment, accurately predicting the amount of water required for cooling can not only optimize the energy efficiency of the equipment but also effectively reduce the waste of water resources.

[0003] In the coal industry, the invention patent application with the patent application number CN202411270400.4 discloses an exhaust gas purification device for a low-temperature coal dry distillation process. This device uses a multi-stage purification system to remove dust, sulfides, and nitrogen oxides in the exhaust gas to ensure that the exhaust gas generated during the coal dry distillation process meets environmental protection standards. However, the structure of this device is relatively complex, and the multi-stage purification system requires a large amount of energy consumption, which affects the economy and sustainability of the exhaust gas treatment process.

[0004] Another application of exhaust gas treatment technology is in industrial exhaust gas treatment equipment. For example, the invention patent application with the patent application number CN202411719657.3 discloses an exhaust gas treatment device, in which gas purification is carried out through a spray tank, and a spiral upward grid plate and a reagent tank system are used in combination for multi-layer purification of exhaust gas. Although this device effectively filters and purifies the gas, its treatment efficiency is affected by the amount of reagent used and the spraying effect, and the efficiency may decrease when treating exhaust gas with a large flow rate.

[0005] In the field of hydrogen purification, the invention patent application with the patent application number CN202411380524.8 discloses a purification system for industrial by-product hydrogen. This system adsorbs hydrogen and a small amount of impurity gases by using a magnesium-based solid hydrogen storage material to achieve the purification effect of hydrogen. The advantage of this device is that the material can be recycled, thereby reducing the waste of resources during the purification process of industrial by-product hydrogen. However, there may still be problems with incomplete treatment effects in the purification of high-concentration impurity gases.

[0006] In the field of high-temperature flue gas purification, a utility model patent with the patent application number CN202323568701.9 discloses a high-temperature flue gas washing and heat removal experimental platform and system. This platform cools and dust-removes high-temperature flue gas through a multi-stage wet chord grid water film dust removal device, and combines a desulfurization and denitrification device for flue gas purification. Although this system can effectively reduce pollutants in high-temperature flue gas, the stability and continuous working ability of the equipment are still a challenge in high-temperature environments.

[0007] In recent years, with the rapid development of artificial intelligence and deep learning technologies, LSTM (Long Short-Term Memory Network), as a deep learning algorithm for processing time series data, has gradually been applied to fields such as temperature prediction and process control. LSTM has powerful time series modeling capabilities and can accurately predict future trends based on historical data. During the flue gas cooling process, LSTM can predict the precise amount of water required for cooling by learning data such as historical flue gas temperature, flow rate, and environmental factors, and dynamically adjust the spray volume according to real-time temperature changes, thereby achieving the best match between water volume and temperature control.

[0008] Currently, although LSTM has achieved remarkable results in many industrial fields, its application in high-temperature flue gas cooling systems is still in its infancy. Most existing research and applications still remain at traditional temperature control methods and lack technologies for intelligent prediction of cooling water consumption. Therefore, developing a method for predicting the water consumption of a flue gas washing and heat removal device based on the LSTM algorithm, which can optimize the water volume usage in the cooling process in real time, not only helps improve the energy efficiency of the system but also reduces water resource waste, and is of great significance for enhancing the overall performance of the flue gas treatment system. Summary of the Invention

[0009] To solve the problems mentioned in the background technology, the present invention proposes a temperature monitoring and regulation method for a flue gas washing and heat removal device based on LSTM.

[0010] The present invention is realized by the following technical solutions:

[0011] In the first aspect, a temperature monitoring and regulation method for a flue gas washing and heat removal device based on LSTM includes:

[0012] S1, the flue gas washing and heat removal device initially operates using the heat balance formula to calculate the ideal water flow rate required for cooling;

[0013] S2, based on the sensing devices deployed on the flue gas washing and heat removal device, obtain the original historical data and real-time data of the device during the initial operation of the device. The original historical data includes temperature data and water flow rate data; map the original historical data to a high-dimensional space based on a kernel function; obtain the enhanced features most relevant to temperature in the original historical data mapped to the high-dimensional space based on an attention mechanism;

[0014] S3. Establish an LSTM model based on the data generated during the cooling process. The LSTM model takes the enhanced features most relevant to the water consumption and temperature changes as inputs and outputs local time series.

[0015] S4. Perform weighted fusion on the enhanced features and local time series to obtain a temperature change prediction model. Input the enhanced features and local time features, and output the predicted water consumption value required for cooling.

[0016] In step S1, deploy a temperature sensor at each of the inlet and outlet of the flue gas washing and heat dissipation device to calculate the temperature difference of the water. One ordinary temperature sensor and one infrared temperature sensor are set at the top of the air inlet, and one ordinary temperature sensor and one infrared temperature sensor are set at the bottom of the air inlet. One ordinary temperature sensor and one infrared temperature sensor are set at the top of the air outlet, and one ordinary temperature sensor and one infrared temperature sensor are set at the bottom of the air outlet. Among them, the infrared temperature sensor is used to monitor the areas in the device that are difficult to directly contact. Calculate the temperature at the air inlet and outlet positions using an appropriate weighted average, and the calculation is as follows:

[0017] T in =w 1 ·T 1 +w 2 ·T 2+ w 3 ·T 3 +w 4 ·T 4

[0018] T out =w 5 ·T 5 +w 6 ·T 6+ w 7 ·T 7 +w 8 ·T 8

[0019] Among them, T in is the temperature at the air inlet position, T 1 is the temperature of the ordinary temperature sensor at the bottom of the air inlet, T 2 is the temperature of the infrared temperature sensor at the bottom of the air inlet, T 3 is the temperature of the ordinary temperature sensor at the bottom of the air outlet, T 4 is the temperature of the infrared temperature sensor at the bottom of the air outlet, T out is the temperature at the air inlet position, T 5 is the temperature of the ordinary temperature sensor at the bottom of the air inlet, T 6 is the temperature of the infrared temperature sensor at the bottom of the air inlet, T 7 is the temperature of the ordinary temperature sensor at the bottom of the air outlet, T 8is the temperature of the infrared temperature sensor at the bottom of the air outlet, and w1, w2, w3, w4, w5, w6, w7, and w8 are all weighted averages.

[0020] Calculate the ideal water consumption required for cooling based on the heat balance formula, and the calculation is as follows:

[0021]

[0022] Among them, M 水 is the ideal water consumption, M 气 is the mass flow rate of the waste gas, C 气 is the specific heat capacity of the waste gas, ΔT 气 is the temperature difference of the waste gas, C 水 is the specific heat capacity of water, ΔT 水 is the temperature difference of water.

[0023] In the initial stage of the device operation, due to the lack of sufficient data for LSTM training, the ideal water consumption required for cooling is calculated using the heat balance formula for cooling treatment, and the original historical data is collected during this period.

[0024] In step S2, mapping the original historical data to a high-dimensional space based on the kernel function includes:

[0025] Preprocess the original historical data to obtain the original historical data feature matrix;

[0026] Map the original historical data feature matrix to a high-dimensional space based on the Gaussian kernel to obtain a kernel matrix reflecting the similarity between the original historical data feature vectors, and the calculation is as follows:

[0027]

[0028] Among them, x i is the original historical data feature vector at the i-th time step, x j is the original historical data feature vector at the j-th time step, ||x i -x j || represents the Euclidean distance between two feature vectors, and σ is the bandwidth parameter of the Gaussian kernel function.

[0029] Obtain the enhanced features most relevant to the temperature in the original historical data mapped to a high-dimensional space based on the attention mechanism, including:

[0030] Calculate the attention score of each original data feature vector based on the kernel matrix, and the calculation is as follows:

[0031]

[0032] Among them, s i is the attention score of the original historical data feature vector at the i-th time step, wij is the weight of the similarity between the original historical data feature vector at the i-th time step and the original historical data feature vector at the j-th time step, and T is the total length of the time series;

[0033] Normalize the attention scores based on the softmax function to obtain the weights of the original historical data feature vectors, which are calculated as:

[0034]

[0035] where ɑ i is the weight of the i-th original historical data feature vector;

[0036] Perform weighted synthesis on all original historical data feature vectors to obtain the enhanced feature vector, which is calculated as:

[0037]

[0038] where y KAN is the enhanced feature vector, and the enhanced feature matrix is obtained based on the enhanced feature vector as:

[0039]

[0040] In step S3, establish an LSTM model based on the data generated during the cooling process. The LSTM model inputs the enhanced features most relevant to the water consumption and temperature changes, and the output local time series includes:

[0041] Calculate the output of the forget gate to control the enhanced features to be retained or forgotten in the memory unit of the previous time step, which is calculated as:

[0042] f t = λ(W f · [h t-1 , y KANt + b f )

[0043] where W f is the weight matrix of the forget gate, h t-1 is the hidden layer state of the previous moment, y KANt is the input enhanced feature of the current time step, b f is the bias term of the forget gate, λ is the sigmoid activation function, and f t is the output of the forget gate at the current time step:

[0044] Calculate the input gates to determine the degree to which the enhanced features of the current time need to be written into the memory unit, which is calculated as:

[0045] i t = λ(W i · [ht-1 , y KANt + b i )

[0046]

[0047] Among them, W i is the weight matrix of the input gate, b i is the bias term of the input gate, W c is the weight matrix for generating candidate values, b c is the corresponding bias term for generating candidate values, is the new candidate value at the current time step, i t is the output of the input gate at the current time step;

[0048] The update of the memory cell state at the current time step is based on the memory cell state at the previous time step and the candidate memory cell of the current input The memory cell state at the current time step is calculated as:

[0049]

[0050] Among them, c t is the memory cell state at the current time step, c t-1 is the memory cell state at the previous time step;

[0051] Calculate the output of the output gate to obtain the hidden layer state at the current time step, which is calculated as:

[0052] o t = λ(W o · [h t-1 , y KANt + b o )

[0053]

[0054] Among them, w o is the weight matrix of the output gate, b o is the bias term of the output gate, o t is the output of the output gate at the current time step, h t is the hidden layer state at the current time step;

[0055] Based on the hidden layer states of all time steps, local time series features are obtained.

[0056] In step S4, the enhanced features and local time series features are weighted and combined, which is calculated as:

[0057] F = β · Y KAN + (1 - β) · H LSTM

[0058] Among them, β is the weight parameter, Y KAN is the enhanced feature matrix, H LSTM is the local time series feature, and F is the feature vector after weighted synthesis;

[0059] Based on the fully connected layer, the feature vector after weighted synthesis is mapped to a low-dimensional space, and the calculation is as follows:

[0060] F final = ReLU(W F ·F + b F )

[0061] Among them, W F is the weight matrix of the fully connected layer, b F is the bias of the fully connected layer, and F final is the output of the fully connected layer;

[0062] Based on the output layer, the output of the fully connected layer is combined with the weight parameter corresponding to the cooling rate to obtain the prediction of the water consumption required for cooling.

[0063] The technical effect of the present invention is that: this method obtains the original historical data through the sensing devices deployed on the flue gas washing and heat dissipation device, and uses the kernel function mapping and attention mechanism to effectively mine the enhanced features most relevant to the water consumption and temperature change, improving the data value. The LSTM model established based on the temperature state time dependence relationship can accurately capture the temporal changes of the features, and the output local time series features are more targeted. The weighted fusion of the enhanced features and the local time series features constructs a prediction model for the water consumption required for cooling, improving the accuracy and reliability of the temperature monitoring of the flue gas washing and heat dissipation device, which has important significance in the field of environmental protection technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a schematic flowchart of the temperature monitoring and control method for the flue gas washing and heat dissipation device based on LSTM according to the embodiment of the present invention.

[0065] Figure 2 is a schematic diagram of the deployment of the temperature sensor of the flue gas washing and heat dissipation device in the temperature monitoring and control method for the flue gas washing and heat dissipation device based on LSTM according to the embodiment of the present invention.

[0066] Figure 3 is a scatter plot of the temperature monitoring data at the air inlet taking the waste gas with an initial temperature of 100 °C as an example in the temperature monitoring and control method for the flue gas washing and heat dissipation device based on LSTM according to the embodiment of the present invention.

[0067] Figure 4 is a scatter plot of the temperature monitoring data at the air outlet of the cooled gas with a target temperature of 30 °C in the temperature monitoring and control method for the flue gas washing and heat dissipation device based on LSTM according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0068] To make the objectives, technical solutions and effects of the present invention clearer and more explicit, the following further describes the present invention in detail with reference to the accompanying drawings and by way of specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0069] In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0070] Please refer to Figure 1 、 Figure 2 , the temperature monitoring and control method of the cigarette washing and heat elimination device based on LSTM in the embodiment of the present invention includes:

[0071] S1, at the initial stage, the cigarette washing and heat elimination device operates by calculating the ideal water flow required for temperature reduction using the heat balance formula;

[0072] Deploy ordinary temperature sensors at position 1 at the bottom of the air inlet of the cigarette washing and heat elimination device, infrared temperature sensors at position 2, ordinary temperature sensors at position 3 at the top of the air inlet, infrared temperature sensors at position 4 each, ordinary temperature sensors at position 5 at the bottom of the air outlet, infrared temperature sensors at position 6 each, ordinary temperature sensors at position 7 at the top of the air outlet, and infrared temperature sensors at position 8 each. Among them, the infrared temperature sensors are used to monitor the areas in the device that are difficult to directly contact. The sensor data at different positions are weighted and averaged to obtain the overall temperature inside the device. This method can adjust the weights according to the accuracy, importance or distance of the sensors. The calculation is as follows:

[0073] T in = w 1 ·T 1 + w 2 ·T 2+ w 3 ·T 3 + w 4 ·T 4

[0074] T out = w 5 ·T 5 + w 6 ·T 6+ w 7 ·T 7 + w 8·T 8

[0075] For different characteristics of the cooling devices, the cooling efficiency is inconsistent. According to different requirements, the water temperature difference can be changed. Based on the heat balance formula, the ideal water consumption required for cooling is calculated as follows:

[0076]

[0077] Where, M 气 is the mass flow rate of the waste gas, C 气 is the specific heat capacity of the waste gas, ΔT 气 is the temperature difference of the waste gas, C 水 is the specific heat capacity of water, ΔT 水 is the temperature difference of water.

[0078] S2. Based on the sensing devices deployed on the flue gas washing and heat elimination device, obtain the original historical data and the real-time data of the device at the initial stage of operation. The original historical data includes temperature data and water flow rate data; map the original historical data to a high-dimensional space based on the kernel function; based on the attention mechanism, obtain the enhanced features most relevant to the temperature in the original historical data mapped to the high-dimensional space;

[0079] The real-time acquired data ensures the timeliness of information, enabling subsequent analysis to promptly capture the dynamic changes in the temperature situation.

[0080] S3. Establish an LSTM model based on the data generated during the cooling process. The LSTM model inputs the enhanced features most relevant to the water consumption and temperature change, and outputs a local time series;

[0081] The LSTM model can effectively learn the law of the cooling state changing over time, and accurately capture the dependency relationship of features between different time steps. The output local time series features not only contain the information of the current moment, but also contain the key trends in the historical data, providing a dynamic and forward-looking basis for predicting the future temperature state, making the prediction results more consistent with the evolution of the actual cooling process, greatly improving the reliability and accuracy of the prediction of the cooling state, and helping to formulate reasonable strategies in advance.

[0082] S4. Perform weighted fusion on the enhanced features and the local time series to obtain a temperature change prediction model. Input the enhanced features and local time features, and output the predicted value of the water consumption required for cooling;

[0083] It can reasonably allocate weights according to the importance of different features, enabling the model to fully absorb the key information in the enhanced features and the temporal dynamic information in the local time series features.

[0084] In the embodiment of the present invention, in step S2, mapping the original historical data to a high-dimensional space based on the kernel function includes:

[0085] Preprocess the original historical data to obtain the original historical data feature matrix;

[0086] Map the original historical data feature matrix to a high-dimensional space based on the Gaussian kernel to obtain a kernel matrix reflecting the similarity between the original historical data feature vectors, calculated as:

[0087]

[0088] where x i is the original historical data feature vector at the i-th time step, x j is the original historical data feature vector at the j-th time step, ||x i -x j || represents the Euclidean distance between two feature vectors, and σ is the bandwidth parameter of the Gaussian kernel function.

[0089] Obtain the enhanced features most relevant to temperature in the original historical data mapped to the high-dimensional space based on the attention mechanism, including:

[0090] Calculate the attention score for each original data feature vector based on the kernel matrix, calculated as:

[0091]

[0092] where, s i is the attention score of the original historical data feature vector at the i-th time step, w ij is the weight of the similarity between the original historical data feature vector at the i-th time step and the original historical data feature vector at the j-th time step, and T is the total length of the time series;

[0093] Normalize the attention scores based on the softmax function to obtain the weights of the original historical data feature vectors, calculated as:

[0094]

[0095] where, ɑ i is the weight of the i-th original historical data feature vector;

[0096] Perform weighted synthesis on all original historical data feature vectors to obtain the enhanced feature vector, calculated as:

[0097]

[0098] where, y KAN is the enhanced feature vector, and the enhanced feature matrix obtained based on the enhanced feature vector is:

[0099]

[0100] In step S3, calculate the output of the forget gate to control the reinforcement features to be retained or forgotten in the memory cell of the previous time step, which is calculated as:

[0101] f t = λ(W f · [h t-1 , y KANt + b f )

[0102] where W f is the weight matrix of the forget gate, h t-1 is the hidden layer state of the previous moment, y KANt is the input reinforcement feature of the current time step, b f is the bias term of the forget gate, λ is the sigmoid activation function, and f t is the output of the forget gate at the current time step:

[0103] Calculate the input gate to determine the degree to which the reinforcement feature of the current time needs to be written into the memory cell, which is calculated as:

[0104] i t = λ(W i · [h t-1 , y KANt + b i )

[0105]

[0106] where W i is the weight matrix of the input gate, b i is the bias term of the input gate, W c is the weight matrix for generating candidate values, b c is the corresponding bias term for generating candidate values, is the new candidate value at the current moment, and i t is the output of the input gate at the current time step;

[0107] The update of the memory cell state at the current time step is based on the memory cell state of the previous time step and the current input candidate memory cell The memory cell state at the current time step is calculated as:

[0108]

[0109] where c t is the memory cell state at the current time step, and c t-1 is the memory cell state of the previous time step;

[0110] Calculate the output of the output gate to obtain the hidden state at the current time step, which is calculated as:

[0111] o t = λ(W o · [h t-1 , y KANt + b o )

[0112]

[0113] where, w o is the weight matrix of the output gate, b o is the bias term of the output gate, o t is the output of the output gate at the current time step, h t is the hidden state at the current time step;

[0114] Based on the hidden states at all time steps, local time series features are obtained.

[0115] In step S4, the enhanced features and local time series features are weighted and synthesized, which is calculated as:

[0116] F = β · Y KAN + (1 - β) · H LSTM

[0117] where, β is the weight parameter, Y KAN is the enhanced feature matrix, H LSTM is the local time series feature, and F is the feature vector after weighted synthesis;

[0118] Based on the fully connected layer, the feature vector after weighted synthesis is mapped to a low-dimensional space, which is calculated as:

[0119] F final = ReLU(W F · F + b F )

[0120] where, W F is the weight matrix of the fully connected layer, b F is the bias of the fully connected layer, and F final is the output of the fully connected layer;

[0121] Based on the output layer, the output of the fully connected layer is combined with the weight parameter corresponding to the cooling rate to obtain the prediction of the water consumption required for cooling.

[0122] Figure 3 This is the scatter plot of the temperature monitoring data at the air inlet in the temperature monitoring and regulation method of the flue gas heat dissipation reduction device based on LSTM according to the embodiment of the present invention, taking the waste gas with an initial temperature of 100°C as an example. According to Figure 3It can be seen that the predicted value and the actual measured value of the temperature at the air inlet of the flue gas washing and heat dissipation device have very consistent change trends, which can well reflect the temperature change at the air inlet of the device and predict the amount of water required to cool down to the target temperature.

[0123] Figure 4 This is a scatter plot of the temperature monitoring data at the outlet of the cooled gas with a target temperature of 30°C in the temperature monitoring and control method of the flue gas washing and heat dissipation device based on LSTM according to the embodiments of the present invention. Figure 4 It can be seen that the measured value of the outlet temperature based on the temperature monitoring and control method of the flue gas washing and heat dissipation device based on LSTM is closer to the set target cooling temperature.

[0124] As described above, an implementation manner is provided in combination with specific content. It is not considered that the specific implementation of the present invention is only limited to these descriptions. At the same time, due to different industry names, it is not limited to the above names or English names. Any method, structure, etc. similar or identical to the present invention, or any technical deduction or replacement made under the premise of the inventive concept of the present invention, should be regarded as within the protection scope of the present invention.

Claims

1. A temperature monitoring and control method for a smoke washing and heat removal device based on LSTM, characterized in that: include: S1, the smoke washing and heat removal device initially uses the heat balance formula to calculate the ideal water flow required for cooling; S2, based on the sensor equipment deployed on the smoke washing and heat removal device, obtains the original historical data and real-time data of the device in the initial operation of the device, and the original historical data includes temperature data and water flow data; Mapping the original historical data to a high-dimensional space based on the kernel function; Based on the attention mechanism, the most enhanced features related to temperature are obtained from the original historical data after being mapped to the high-dimensional space; S3, build an LSTM model based on the data generated during the cooling process. The LSTM model inputs the most relevant reinforcement features of water consumption and temperature changes and outputs a local time series; S4, weighted fusion of enhanced features and local time series is performed to obtain a temperature change prediction model, which inputs enhanced features and local time features and outputs the predicted value of water consumption required for cooling.

2. The temperature monitoring and control method of the smoke washing and heat removal device based on LSTM according to claim 1 is characterized in that: The smoke washing and heat removal device initially uses a heat balance formula to calculate the ideal water flow required for cooling, and the operation includes: The temperature at the inlet and outlet is calculated using an appropriate weighted average: T in =w1·T1+w2·T 2+ w3·T3+w4·T4 T out =w5·T5+w6·T 6+ w7·T7+w8·T8 Among them, T in is the temperature at the air inlet, T1 is the temperature of the common temperature sensor at the bottom of the air inlet, T2 is the temperature of the infrared temperature sensor at the bottom of the air inlet, T3 is the temperature of the common temperature sensor at the bottom of the air outlet, T4 is the temperature of the infrared temperature sensor at the bottom of the air outlet, T out is the temperature at the air inlet, T5 is the temperature of the common temperature sensor at the bottom of the air inlet, T6 is the temperature of the infrared temperature sensor at the bottom of the air inlet, T7 is the temperature of the common temperature sensor at the bottom of the air outlet, T8 is the temperature of the infrared temperature sensor at the bottom of the air outlet, w1, w2, w3, w4, w5, w6, w7, w8 are all weighted averages; The ideal water consumption required for cooling is calculated based on the heat balance formula: Among them, M 水 is the ideal water consumption, M 气 is the mass flow rate of exhaust gas, C 气 is the specific heat capacity of the exhaust gas, ΔT 气 is the temperature difference of the exhaust gas, C 水 is the specific heat capacity of water, ΔT 水 is the temperature difference of water.

3. The temperature monitoring and control method of the smoke washing and heat removal device based on LSTM according to claim 1 is characterized in that: In step S2, mapping the original historical data to a high-dimensional space based on the kernel function includes: Preprocess the original historical data to obtain the original historical data feature matrix; Based on the Gaussian kernel, the original historical data feature matrix is ​​mapped to a high-dimensional space to obtain a kernel matrix that reflects the similarity between the feature vectors of the original historical data, which is calculated as: where x i is the original historical data feature vector of the i-th time step, x j is the feature vector of the original historical data at the jth time step, ||x i -x j || represents the Euclidean distance between two eigenvectors, and σ is the bandwidth parameter of the Gaussian kernel function.

4. The temperature monitoring and control method of the smoke washing and heat removal device based on LSTM according to claim 3 is characterized in that: In step S2, the most temperature-related enhanced features in the original historical data mapped to the high-dimensional space are obtained based on the attention mechanism, including: The attention score of each original data feature vector is calculated based on the kernel matrix, which is calculated as: Among them, s i is the attention score of the original historical data feature vector at the i-th time step, w ij is the weight of the similarity between the feature vector of the original historical data at the i-th time step and the feature vector of the original historical data at the j-th time step, and T is the total length of the time series; The attention scores are normalized based on the softmax function to obtain the weight of the original historical data feature vector, which is calculated as: Among them, i is the weight of the i-th original historical data feature vector; All original historical data feature vectors are weighted and integrated to obtain the enhanced feature vector, which is calculated as: Among them, y KAN To strengthen the feature vector, the strengthened feature matrix is ​​obtained based on the strengthened feature vector:

5. The temperature monitoring and control method of the smoke washing and heat removal device based on LSTM according to claim 4 is characterized in that: In step S3, an LSTM model is established based on the data generated during the cooling process. The LSTM model inputs the most relevant reinforcement features of water consumption and temperature changes, and outputs local time series including: The forget gate output is calculated to control the reinforcement features that need to be retained or forgotten in the memory unit of the previous time step, which is calculated as: f t =λ(W f ·[h t-1 ,y KANt ]+b f ) Among them, W f is the weight matrix of the forget gate, h t-1 is the hidden state of the previous moment, y KANt is the input enhancement feature of the current time step, b f is the bias term of the forget gate, λ is the sigmoid activation function, and ft is the output of the forget gate at the current time step: Calculate the inputs to determine the extent to which the reinforcement features of the current time need to be written into the memory cell, calculated as: i t =λ(W i ·[h t-1 ,y KANt ]+b i ) Among them, W i is the weight matrix of the input gate, b i is the bias term of the input gate, W c is the weight matrix used to generate candidate values, b c is the corresponding bias term used to generate candidate values, is the new candidate value at the current moment, i t The output of the input gate for the current time step; The update of the memory cell state at the current time step is based on the memory cell state at the previous time step and the candidate memory cell c currently input. t , the memory cell state at the current time step is calculated as: Among them, c t is the memory cell state at the current time step, c t-1 is the memory cell state at the previous time step; Calculate the output gate output to get the hidden state of the current time step, calculated as: o t =λ(W o ·[h t-1 ,y KANt ]+b o ) Among them, w o is the weight matrix of the output gate, b o is the bias term of the output gate, o t is the output of the output gate at the current time step, h t is the hidden layer state of the current time step; Based on the hidden layer states of all time steps, local time series features are obtained.

6. The temperature monitoring and control method of the smoke washing and heat removal device based on LSTM according to claim 5 is characterized in that: In step S4, the enhanced features and local time series features are weighted and integrated, and calculated as: F=β·Y KAN +(1-β)·H LSTM Among them, β is the weight parameter, Y KAN is the enhanced feature matrix, H LSTM is the local time series feature, and F is the feature vector after weighted integration; Based on the fully connected layer, the weighted integrated feature vector is mapped to a low-dimensional space and calculated as: F final =ReLU(W F ·F+b F ) Among them, W F is the weight matrix of the fully connected layer, b F is the bias of the fully connected layer, F final is the output of the fully connected layer; Based on the output layer, the output of the fully connected layer is combined with the weight parameter corresponding to the cooling rate to obtain the water consumption prediction required for cooling.

Images