A combustion state prediction method based on combination of random forest and LSTM

Abstract

The present disclosure provides a combustion state prediction method based on the combination of random forest and LSTM, by constructing a cascaded prediction framework, first using long short-term memory network to extract deep features of time series data of combustion process, capture its internal dynamic change law and long-term dependence, and then input the extracted high-level features into the random forest model for final state discrimination and prediction. This method fully utilizes the advantages of LSTM in processing time series data and the ability of random forest in robust classification in high-dimensional feature space, through the synergy and complementation of the two, effectively overcomes the shortcomings of traditional single method in prediction accuracy, generalization ability and modeling of complex nonlinear relationship, so as to realize more accurate and reliable prediction of combustion state.

Description

Technical Field

[0001] This invention relates to the field of industrial water treatment technology, and in particular to a combustion state prediction method based on a combination of random forest and LSTM. Background Technology

[0002] Combustion plays a vital role in basic industries such as energy, power, chemical engineering, and metallurgy, as well as high-tech fields like aerospace. The quality of combustion directly affects the normal operation of various production activities. Many factors actually influence combustion state, and the characteristic space of solutions to problems aimed at determining and predicting combustion state has a very high number of bits, making it difficult to guarantee the speed and accuracy of general computational methods. For example, traditional neural network learning algorithms improve the accuracy of solutions by increasing the algorithm's scale, but this also increases computational complexity and practical computational risk. Furthermore, this algorithm uses the principle of empirical risk minimization, which cannot minimize expected risk, thus having theoretical flaws.

[0003] Neural network learning algorithms have become one of the most commonly used methods for combustion state prediction due to their advantages such as strong data fitting ability, strong generalization ability, high adaptability and flexibility, and ease of combination with other machine learning algorithms. In recent years, Long Short-Term Memory (LSTM) networks have been widely used in combustion prediction. Compared with traditional machine learning algorithms, combining LSTM networks with other machine learning algorithms such as random forests improves the accuracy of data prediction. Summary of the Invention

[0004] The first aspect of this disclosure provides a method for predicting combustion state based on a combination of random forest and LSTM, comprising the following steps: S1: Establish a long short-term memory network model and determine its network parameters; S2: Use the Long Short-Term Memory Network model to process the original time series data, extract the nonlinear dynamic features of the time series, and obtain the hidden state or predicted value of the time series. S3: Establish a random forest algorithm model and determine its model parameters; S4: Input the hidden state or predicted value output by the Long Short-Term Memory Network model into the Random Forest Algorithm model to obtain the final burning state prediction result.

[0005] In conjunction with the first aspect, in step S2, before processing the original time series data using the Long Short-Term Memory Network model, the step further includes a data cleaning and preprocessing step for the original time series data. The data cleaning and preprocessing includes at least one of deleting duplicate data, repairing erroneous data, filling in missing values, handling outliers, and data format conversion.

[0006] In conjunction with the first aspect, the missing values ​​are filled using interpolation or mean imputation.

[0007] In conjunction with the first aspect, in step S1, the parameters of the long short-term memory network model include the time window length, the number of hidden layer units, the learning rate, and the discard rate.

[0008] In conjunction with the first aspect, in step S2, the long short-term memory network model models the trend and seasonality information in the time series data through its included forget gate, input gate, cell state update unit, and output gate.

[0009] In conjunction with the first aspect, in step S3, the random forest algorithm model consists of multiple classification and regression trees, and the model parameters include the number of binary trees and the maximum depth.

[0010] In conjunction with the first aspect, the process of constructing the classification and regression tree includes: S31: Multiple training set copies are generated from the original sample set by sampling with replacement, and each copy is used to train a classification and regression tree; S32: For each copy of the training set, recursively perform the following operations to construct the binary tree: Iterate through each feature in the current feature set and calculate the optimal split point and its corresponding Gini coefficient for each feature. Select the feature with the smallest Gini coefficient and its optimal split point as the current node, and divide the current sample set into two subsets based on the split point; For each subset obtained by partitioning, update the feature set and return to step a) to recursively build the tree until the preset termination condition is met.

[0011] In conjunction with the first aspect, in step S4, the prediction results output by the Long Short-Term Memory Network model are weighted and averaged before being input into the Random Forest Algorithm Model. The original time series data is the historical operating pressure pulsation data of the gas turbine.

[0012] A second aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the combustion state prediction method.

[0013] A third aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, enables the method for classifying and diagnosing faults in a thermal power plant switchgear.

[0014] Beneficial Effects: This disclosure presents a combustion state prediction method based on a combination of random forest and LSTM. By constructing a cascaded prediction framework, it first utilizes a Long Short-Term Memory (LSTM) network to extract deep features from the time-series data of the combustion process, capturing its inherent dynamic changes and long-term dependencies. Then, the extracted high-level features are input into a random forest model for final state discrimination and prediction. This method fully leverages the advantages of LSTM in processing time-series data and the robust classification ability of random forest in high-dimensional feature spaces. Through the synergy and complementarity of the two, it effectively overcomes the shortcomings of traditional single methods in prediction accuracy, generalization ability, and modeling of complex nonlinear relationships, thereby achieving more accurate and reliable prediction of combustion state. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a combustion state prediction method based on a combination of random forest and LSTM according to an embodiment of this disclosure. Figure 2 This is a time series diagram comparing the predicted values ​​and actual values ​​of the random forest in this embodiment of the present disclosure. Figure 3 This is a time series diagram comparing the predicted and actual values ​​of the random forest and LSTM combination prediction in an embodiment of this disclosure. Figure 4 An electronic device according to an embodiment of this disclosure. Detailed Implementation

[0016] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those disclosed herein.

[0017] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the present disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0018] Figure 1 This is a flowchart illustrating a combustion state prediction method based on a combination of random forest and LSTM, according to an embodiment of this disclosure, including: S1: Establish a long short-term memory network model and determine its network parameters; In step S1, the parameters of the long short-term memory network model include the time window length, the number of hidden layer units, the learning rate, and the dropout rate.

[0019] Specifically, Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network (RNN) that can effectively capture long-term dependencies. LSTM models trends and seasonality information in time series data through gating mechanisms (input gate, forget gate, output gate) and cell states.

[0020] The time window length defines the number of historical data points observed by the model in each prediction. Its value needs to be determined based on the periodicity of the dynamic changes in the combustion process or key time scales. For example, a length that can effectively capture key antecedent states can be selected by analyzing the autocorrelation function of historical pressure pulsation data or based on domain knowledge (such as the characteristic times of specific combustion instability modes). In a preferred embodiment, the model's prediction performance (e.g., root mean square error) on the validation set is compared experimentally with different window lengths (e.g., 10, 20, 50, and 100 time steps), and the window length that optimizes performance is ultimately selected.

[0021] The number of hidden layer units determines the dimension of the internal state vector of the LSTM network, i.e., the model's capacity and ability to extract complex features. Too few units may lead to underfitting, failing to fully learn the complex patterns of the data; too many units may lead to overfitting and increase computational burden. Grid search or empirical methods are typically used to set the number of units. In one embodiment, starting with a set of preset values ​​(e.g., 32, 64, 128, 256), the model is trained on the training set and evaluated on an independent validation set. The number of units that achieves the best prediction accuracy on the validation set without significant overfitting is selected as the final parameter.

[0022] The learning rate controls the step size of model parameter updates in each iteration and is a key hyperparameter affecting training convergence speed and final performance. An excessively high learning rate may lead to training instability or even divergence, while an excessively low learning rate will slow down the convergence process. Adaptive optimization algorithms (such as Adam) are typically used, with an initial learning rate set. This initial value can be set empirically (e.g., 0.001, 0.01) or adjusted by observing the loss function's decline curve through small-scale experiments. In one embodiment, a learning rate decay strategy is employed; for example, during training, when the validation set loss no longer decreases, the learning rate is halved to promote model convergence to a better local minimum.

[0023] The dropout rate refers to the proportion of hidden layer units that are randomly "dropped out" (i.e., temporarily ignored) during training. This is an effective regularization technique used to prevent overfitting. The dropout rate is typically chosen between 0 and 0.5. By training the model at different dropout rates (e.g., 0.1, 0.2, 0.3, 0.5) and comparing its performance on the validation set, an optimal value that effectively improves the model's generalization ability can be determined. An excessively high dropout rate may hinder the model's learning ability.

[0024] After determining the aforementioned key parameters, an LSTM network structure is constructed. This network consists of an input layer, one or more LSTM layers (each containing the aforementioned predetermined number of hidden units, with an adjustable dropout rate), and an output layer. The network works in conjunction with the cell state through its internal gating mechanisms (including forget gates, input gates, and output gates) to process the input combustion state time-series data step-by-step. The forget gate determines which information to discard from the cell state; the input gate determines which new information to store in the cell state; the updated cell state carries long-term memory information; and the output gate determines the hidden state output for that time step based on the current input and cell state. Through iterative training, the model learns the complex mapping relationship from historical sequence data to future states or features.

[0025] After completing the model structure construction and parameter initialization, the LSTM model is trained using the prepared training dataset until the model converges, thus obtaining a usable Long Short-Term Memory network model.

[0026] S2: Use the Long Short-Term Memory Network model to process the original time series data, extract the nonlinear dynamic features of the time series, and obtain the hidden state or predicted value of the time series. In step S2, before processing the original time series data using the Long Short-Term Memory Network model, the step further includes data cleaning and preprocessing of the original time series data. The data cleaning and preprocessing includes at least one of deleting duplicate data, repairing erroneous data, filling in missing values, handling outliers, and data format conversion.

[0027] The missing values ​​are filled using interpolation or mean imputation.

[0028] In step S2, the long short-term memory network model models the trend and seasonality information in the time series data through its forgetting gate, input gate, cell state update unit and output gate.

[0029] The specific steps are as follows: Data cleaning and preprocessing: Preprocessing the raw time-series data reflecting the combustion state (e.g., continuous readings from sensors such as pressure, temperature, and flow rate).

[0030] Remove duplicate data: Identify and remove identical duplicate records caused by data acquisition or transmission errors.

[0031] Correcting erroneous data: Based on physical laws or process knowledge, set reasonable thresholds, identify obviously erroneous data points (such as extreme values ​​exceeding the range), and replace them with the mean of the valid data before and after, or interpolation based on trends.

[0032] Imputing missing values: For missing points in data records, interpolation methods (such as linear interpolation or spline interpolation) or statistical values ​​(such as the mean or median) of valid data within a time window before and after the missing point are used to impute them. This step ensures the continuity of the time series and is crucial for subsequent sequence modeling.

[0033] Handling outliers: Use statistical methods (such as those based on standard deviation or interquartile range) to identify and handle outliers. For outliers that are not part of a systematic error, truncation, Windsorization, or treating them as missing values ​​and handling them using the imputation methods described above can be applied, depending on the circumstances.

[0034] Data format conversion: Convert timestamps into a format that the model can process (such as timestamp objects or seconds relative to the start point), and normalize or standardize all feature data to a uniform numerical range (such as [0,1] or a mean of 0 and a variance of 1) to accelerate model training and improve stability.

[0035] Feature extraction and modeling: The cleaned and preprocessed time series data are organized into a continuous sample sequence according to the time window length determined in step S1, and then input into the trained long short-term memory network model.

[0036] When the model is running, its internal forget gate calculates a value between 0 and 1 based on the current input and the hidden state of the previous time step, which determines which historical information to discard from the cell state.

[0037] The input gate computes two parts simultaneously: one part is the sigmoid layer that determines which values ​​need to be updated; the other part is the tanh layer that generates a new candidate value vector to be added to the cell state.

[0038] Subsequently, the cell state is updated: first, the old cell state is multiplied by the output of the forget gate to discard the information to be forgotten; then, the product of the input gate's output and the candidate value vector is added to add new information. This cell state persists throughout the entire time series, carrying long-term dependency information, thereby effectively capturing trend changes during the combustion process.

[0039] Finally, the output gate determines the hidden state output for the current time step based on the current input, the hidden state from the previous time step, and the updated cell state. This hidden state contains a comprehensive representation of all nonlinear dynamic features related to the target prediction extracted by the model from the historical sequence up to the current time step.

[0040] Through the synergistic effect of the above gating mechanisms, the LSTM model can adaptively learn and memorize long-term patterns (trends) and periodic patterns (seasonality) in the combustion process time series, and suppress interference from irrelevant noise.

[0041] Output results: The output of step S2 has two main application modes: Mode 1 (Feature Extraction): The hidden state (ht) of the last time step corresponding to each time window is used as the output. This hidden state is a high-order feature vector after deep processing by LSTM, representing the condensed information of this historical sequence, and will be used as the input feature of the subsequent random forest model.

[0042] Mode 2 (Preliminary Prediction): A fully connected layer is added as the output layer at the end of the LSTM model, enabling the model to directly output predicted values ​​for the combustion state at the next one or more time steps. These predicted values ​​already contain the temporal patterns learned by the LSTM and can be further refined or integrated by subsequent random forest models.

[0043] Through step S2, the original low-order, high-noise time-series data is transformed into high-quality, high-information-content feature representations or preliminary predictions, laying a solid foundation for subsequent accurate state discrimination based on random forests.

[0044] S3: Establish a random forest algorithm model and determine its model parameters; In step S3, the random forest algorithm model consists of multiple classification and regression trees, and the model parameters include the number of binary trees and the maximum depth.

[0045] The process of constructing the classification and regression tree includes: S31: Multiple training set copies are generated from the original sample set by sampling with replacement, and each copy is used to train a classification and regression tree; S32: For each copy of the training set, recursively perform the following operations to construct the binary tree: Iterate through each feature in the current feature set and calculate the optimal split point and its corresponding Gini coefficient for each feature. Select the feature with the smallest Gini coefficient and its optimal split point as the current node, and divide the current sample set into two subsets based on the split point; For each subset obtained by partitioning, update the feature set and return to step a) to recursively build the tree until the preset termination condition is met.

[0046] Specifically, after completing feature extraction based on the Long Short-Term Memory network, step S3 is executed: establish a random forest algorithm model and determine its model parameters.

[0047] Random forest is an ensemble learning model whose core consists of multiple independent classification and regression trees with potentially different structures. The key to building this model lies in determining its hyperparameters and constructing each CART tree.

[0048] Model parameters determined: The number of binary trees: This refers to the number of decision trees in the forest. Too few trees may not fully leverage the advantages of ensemble learning, while too many trees will increase computational costs and may lead to diminishing returns. Typically, cross-validation is used to select trees within a preset range (e.g., 50, 100, 200, 500) to strike a balance between prediction accuracy and computational efficiency.

[0049] Maximum tree depth: Limiting the maximum depth of each decision tree is an important regularization parameter to prevent model overfitting. It can be set to no limit (until the node purity is too high or the number of samples is too small), or adjusted to a fixed value (e.g., 10, 20, no limit) based on the validation set performance.

[0050] Constructing classification and regression trees: In a random forest, the training set for each CART tree is generated through Bootstrap sampling (sampling with replacement). Specifically, if the original training sample set is N in size, it is randomly sampled N times with replacement to form a "bag" of samples of the same size N. This sample set serves as the basis for training a single CART tree, and this process is repeated until a specified number of training set copies are generated.

[0051] For each copy of the training set obtained through Bootstrap sampling, recursively execute the following algorithm to construct a binary decision tree (CART): Optimal split point selection: Iterate through each feature in the set of candidate features for the current node.

[0052] For the current feature, determine all possible candidate split points based on its feature type (continuous or discrete).

[0053] For each candidate split point, the sample set of the current node is divided into two subsets, left and right (for continuous features, this is usually "feature value ≤ split point" and "feature value > split point"; for discrete features, it is "belongs to a subset" and "does not belong to a subset").

[0054] The Gini coefficient is used as a measure of impurity. The Gini coefficients of the left and right subsets after the split are calculated separately, and the weighted split Gini coefficient corresponding to the split point is calculated with their sample size as the weight.

[0055] For this feature, the candidate point that minimizes the weighted Gini coefficient is selected as its optimal segmentation point, and this minimum Gini coefficient value is recorded.

[0056] Node generation and data partitioning: After traversing all candidate features, compare the minimum Gini coefficient corresponding to each feature.

[0057] The feature with the smallest Gini coefficient and its corresponding optimal split point are selected as the splitting rule for the current node, and an internal node is generated.

[0058] Based on this rule, the sample set of the current node is precisely divided into two subsets, which are then assigned to the left and right branches of the node, respectively.

[0059] Recursive tree construction and termination condition: For each subset obtained (i.e. each newly generated child node), treat it as the new "current node".

[0060] Remove features that have already been used for splitting from the feature set (or decide whether to reuse them based on the strategy), and update the candidate feature set.

[0061] On the new current node, repeat the above steps to recursively split and build subtrees.

[0062] The recursive process terminates at the current node (which becomes a leaf node) when any of the following preset termination conditions are met: The current node depth has reached the preset "maximum depth".

[0063] The current node contains fewer samples than the preset minimum.

[0064] The target variable (combustion state category) of all samples in the current node is completely consistent (Gini coefficient is 0).

[0065] All features have been used for splitting, or no valid split point can be found.

[0066] Model ensemble: Repeat the above process to build a specified number of CART trees (e.g., 100), which together form the final random forest model. During the prediction phase, for a new input sample (i.e., features extracted by LSTM or preliminary predictions), each CART tree independently outputs a prediction result (burning state category). The final prediction output of the random forest is determined by the prediction results of all trees through a voting method (classification task) or an averaging method (regression task), thus integrating the judgments of numerous weak learners to obtain a more stable and accurate strong prediction model.

[0067] The random forest model established in step S3 has a strong ability to handle high-dimensional features, nonlinear relationships, and resist overfitting, thus preparing for the final accurate prediction of combustion state.

[0068] S4: Input the hidden state or predicted value output by the Long Short-Term Memory Network model into the Random Forest Algorithm model to obtain the final burning state prediction result.

[0069] In step S4, the prediction results output by the Long Short-Term Memory Network model are weighted and averaged, and then input into the Random Forest Algorithm Model. The original time series data is the historical operating pressure pulsation data of the gas turbine.

[0070] Specifically, after completing the feature extraction of the Long Short-Term Memory network and the construction of the Random Forest model, the hidden state or predicted value output by the Long Short-Term Memory network model is input into the Random Forest algorithm model to perform the final ensemble prediction and obtain the final combustion state discrimination or prediction result.

[0071] This step embodies the core idea of ​​the cascaded hybrid model proposed in this invention, and the specific implementation process is as follows: Model input preparation: The output obtained by the Long Short-Term Memory Network in step S2 after processing the original time series is used as the input features of the Random Forest model constructed in step S3.

[0072] If feature extraction mode is used, the input is the hidden state vector (ht) of the last time step of the LSTM model. This vector is a high-order abstract representation of the original time series data after deep nonlinear transformation.

[0073] If the preliminary prediction mode is used, the input is the sequence of predicted values ​​for the target variable directly output by the LSTM model. This sequence already contains the temporal dynamics learned by the LSTM.

[0074] In a preferred embodiment, to further smooth the LSTM output and integrate information from multiple time steps, thereby improving the robustness of the input features, a weighted average can be applied to the multiple predicted values ​​output by the LSTM model at adjacent time windows or different time steps. For example, a moving average or an exponentially weighted average that assigns higher weights to recent predicted values ​​can be used to form a more representative scalar or vector feature, which can then be used as the input to the random forest.

[0075] Ensemble prediction execution: Input the prepared input feature matrix (each row corresponds to a sample, and each column corresponds to an LSTM output feature) into the trained random forest model.

[0076] In a random forest model, each classification and regression tree independently performs forward propagation on the input samples. Based on the internal structure of each tree, starting from the root node, the sample is directed to the left or right subtree according to the splitting rules (features and thresholds) on the node, until a leaf node is reached.

[0077] Each leaf node of the CART tree provides a prediction of the burning state (e.g., a specific state category label in a classification task, or a specific numerical value in a regression task).

[0078] The random forest model aggregates the individual predictions from all trees. For classification tasks (such as determining normal / abnormal combustion status), a voting method is used: the category labels given by all trees are counted, and the category with the most votes is taken as the final prediction result of the random forest. For regression tasks (such as predicting specific pressure pulsation amplitudes), an averaging method is used: the arithmetic mean of the output predictions of all trees is calculated as the final prediction result.

[0079] Output and Validation: The final result output by the random forest model is the prediction of the combustion state at a specific moment or time period by the method of this invention. In a specific application embodiment, the original time series data is the historical operating pressure pulsation data of the gas turbine. This method can achieve accurate prediction of the future trend or state of pressure oscillations (reflecting combustion stability) within the gas turbine combustion chamber.

[0080] Figure 2 This is a time series plot comparing the predicted values ​​and actual values ​​of a random forest prediction. Figure 3 A time series plot comparing predicted and actual values ​​for a combination of random forest and LSTM predictions. The horizontal axis represents time, and the vertical axis represents pressure fluctuations. Blue represents the actual value, yellow represents the predicted value of random forest alone, and red represents the predicted value of random forest combined with LSTM.

[0081] This demonstrates that when the proposed combined random forest and LSTM model is used to predict gas turbine pressure pulsation data, the predicted curve shows a significantly higher degree of agreement with the actual observations compared to using only a single random forest model. Furthermore, the prediction error distribution is more concentrated and the amplitude is smaller. This intuitively verifies that the model cascading and integration strategy implemented in step S4 effectively combines the dual advantages of LSTM in temporal feature extraction and random forest in robust decision-making in high-dimensional space, thereby significantly improving the overall accuracy and reliability of combustion state prediction.

[0082] Thus, by sequentially executing steps S1 to S4, the complete combustion state prediction process, from data preparation and in-depth extraction of temporal features to integrated decision-making, has been completed.

[0083] Electronic device 400 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 400 may include, but is not limited to, processor 401 and memory 402. Those skilled in the art will understand that... Figure 4 This is merely an example of electronic device 400 and does not constitute a limitation on electronic device 400. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0084] Processor 401 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0085] The memory 402 can be an internal storage unit of the electronic device 400, such as a hard disk or RAM of the electronic device 400. The memory 402 can also be an external storage device of the electronic device 400, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 400. Furthermore, the memory 402 can include both internal and external storage units of the electronic device 400. The memory 402 is used to store the computer program 403 and other programs and data required by the electronic device. The memory 402 can also be used to temporarily store data that has been output or will be output.

[0086] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit it. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A method for predicting combustion state based on a combination of random forest and LSTM, characterized in that, Includes the following steps: S1: Establish a long short-term memory network model and determine its network parameters; S2: Use the Long Short-Term Memory Network model to process the original time series data, extract the nonlinear dynamic features of the time series, and obtain the hidden state or predicted value of the time series. S3: Establish a random forest algorithm model and determine its model parameters; S4: Input the hidden state or predicted value output by the Long Short-Term Memory Network model into the Random Forest Algorithm model to obtain the final burning state prediction result.

2. The combustion state prediction method according to claim 1, characterized in that, In step S2, before processing the original time series data using the Long Short-Term Memory Network model, the step further includes data cleaning and preprocessing of the original time series data. The data cleaning and preprocessing includes at least one of deleting duplicate data, repairing erroneous data, filling in missing values, handling outliers, and data format conversion.

3. The method according to claim 2, characterized in that, The missing values ​​are filled using interpolation or mean imputation.

4. The method according to claim 1, characterized in that, In step S1, the parameters of the long short-term memory network model include the time window length, the number of hidden layer units, the learning rate, and the dropout rate.

5. The method according to claim 1, characterized in that, In step S2, the long short-term memory network model models the trend and seasonality information in the time series data through its forgetting gate, input gate, cell state update unit and output gate.

6. The method according to claim 1, characterized in that, In step S3, the random forest algorithm model consists of multiple classification and regression trees, and the model parameters include the number of binary trees and the maximum depth.

7. The method according to claim 6, characterized in that, The process of constructing the classification and regression tree includes: S31: Multiple training set copies are generated from the original sample set by sampling with replacement, and each copy is used to train a classification and regression tree; S32: For each copy of the training set, recursively perform the following operations to construct the binary tree: Iterate through each feature in the current feature set and calculate the optimal split point and its corresponding Gini coefficient for each feature. Select the feature with the smallest Gini coefficient and its optimal split point as the current node, and divide the current sample set into two subsets based on the split point; For each subset obtained by partitioning, update the feature set and return to step a) to recursively build the tree until the preset termination condition is met.

8. The method according to claim 1, characterized in that, In step S4, the prediction results output by the Long Short-Term Memory Network model are weighted and averaged, and then input into the Random Forest Algorithm Model. The original time series data is the historical operating pressure pulsation data of the gas turbine.

9. An electronic device, characterized in that, include: One or more processors; A storage unit is used to store one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the fault classification and diagnosis method for thermal power plant switchgear according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it can implement the method for classifying and diagnosing faults in thermal power plant switchgear according to any one of claims 1 to 8.

Images