Thermal power plant operation risk intelligent early warning method based on deep learning

Abstract

The invention discloses a thermal power plant operation risk intelligent early warning method based on deep learning. The method comprises the following steps: S1, collecting operation data to construct a multi-modal time sequence; s2, respectively inputting modal sub-networks after standardization processing; s3, setting a reversible residual structure for each sub-network and introducing a dynamic scaling mechanism; s4, introducing a time gating mechanism to adjust residual response; s5, fusing the modal features to form fusion representation; s6, performing forward transformation, combining with intermediate state reverse reconstruction, and calculating a residual error; s7, a sliding window aggregates residual errors to generate a scoring sequence, and risk levels are output; and S8, generating a key factor sequence based on residual attribution analysis features. According to the invention, operation abnormity in-advance identification and risk source accurate positioning under complex working conditions are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial intelligent monitoring, and particularly relates to a power plant operation risk intelligent early warning method based on deep learning. BACKGROUND

[0002] At present, as a basic energy unit in the power system of China, the safe and stable operation of the power plant is of great significance to the guarantee of power supply and energy scheduling. With the large-scale of thermal power units and the automation of control systems, the data types involved in the operation of the power plant are increasingly complex, covering temperature, pressure, speed, current and control signals and other multi-source working condition variables. Traditional operation risk early warning methods are mostly based on rule-based reasoning, expert systems or shallow machine learning models such as support vector machines, decision trees and random forests, and their core relies on artificially defined feature variables and fixed mode risk identification logic. However, with the increasing non-linear coupling of the operation state and the frequent changes in the dynamic correlation between devices, the traditional methods have gradually exposed their limitations in processing high-dimensional time series data, multi-modal information fusion and fault attribution.

[0003] Some studies attempt to introduce deep neural networks to model the operation data of the power plant, but conventional feedforward neural networks or convolutional networks often face two problems. On the one hand, there is a lack of effective feature reversible modeling mechanism, resulting in uncontrollable information loss after multi-layer processing of input information, affecting the accuracy of risk discrimination; on the other hand, most existing models focus on single modal or unified structure extraction, making it difficult to set exclusive processing paths for different types of operation data (such as sensor values and discrete control signals), causing distortion in modal information fusion. In addition, for residual risk indicators, most existing methods use direct comparison or moving average methods, lack a credibility evaluation mechanism and an explainable analysis path for key factors, which is not conducive to actual fault location and intelligent decision support.

[0004] Therefore, the existing power plant operation risk early warning technology still has the following main defects: first, the modeling capability for multi-modal operation data is insufficient, making it difficult to accurately extract the deep interaction between heterogeneous features; second, there is a lack of reversible feature extraction mechanism, the reconstruction capability is weak, and the information retention degree cannot be quantified; third, there is a lack of time dynamic-oriented residual adjustment and scoring mechanism, and the risk judgment result is not stable and robust; fourth, there is a lack of explainable risk attribution method combined with residual mechanism, which cannot output a reliable key factor ranking result, affecting the reliability of the downstream disposal response.

[0005] Therefore, how to provide a power plant operation risk intelligent early warning method based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0006] One purpose of the present application is to provide a deep learning-based thermal power plant operation risk intelligent early warning method, which combines reversible residual network, dynamic residual scaling mechanism, time gating mechanism and residual attribution algorithm, and describes in detail the whole process of multi-modal operation data modeling, residual reconstruction scoring and key factor inversion analysis, and has the advantages of strong reversible feature extraction, high risk identification accuracy and strong abnormality tracing capability.

[0007] According to the deep learning-based thermal power plant operation risk intelligent early warning method of the embodiment of the present application, the following steps are included:

[0008] S1, collecting operation data in the operation process of the thermal power plant, and constructing a multi-modal time sequence input sequence;

[0009] S2, standardizing the multi-modal time sequence input sequence, and inputting each kind of operation data into the corresponding modal subnetwork;

[0010] S3, setting a reversible residual structure in each modal subnetwork to extract modal features, introducing a dynamic scaling mechanism in the residual path, calculating a scaling factor according to the input data, and adjusting the residual output amplitude;

[0011] S4, introducing a time gating mechanism in the reversible residual structure, calculating a gating coefficient according to the time stamp, and adjusting the residual response of each time step;

[0012] S5, fusing the modal features adjusted by the residual through a reversible mapping structure to form a fused feature representation;

[0013] S6, performing a forward transformation operation on the fused feature representation, recording the intermediate state, and combining the intermediate state to perform reverse reconstruction, calculating the residual value between the original input sequence and the reconstructed sequence;

[0014] S7, constructing a risk scoring module, aggregating the residuals using a sliding time window to generate a score sequence, comparing the score sequence with a preset threshold, and outputting a risk level label;

[0015] S8, setting an attribution module in the reverse reconstruction path, analyzing the input features based on the residual value, and outputting a key factor ranking result.

[0016] Optionally, the operation data includes temperature data, pressure data, speed data, current data and control signal data.

[0017] Optionally, the S2 specifically includes:

[0018] S21, standardizing each kind of operation data in the multi-modal time sequence input sequence respectively;

[0019] S22, input the normalized temperature data into the temperature modal subnetwork, input the pressure data into the pressure modal subnetwork, input the rotating speed data into the rotating speed modal subnetwork, input the current data into the current modal subnetwork, and input the control signal data into the control signal modal subnetwork.

[0020] Optionally, the S3 specifically comprises:

[0021] S31, setting a reversible residual structure in each modal subnetwork, the reversible residual structure comprising a main path and a residual path, the residual path receiving the modal input x t after the normalization processing in step S2 t ;

[0022] S32, embedding a dynamic scaling mechanism in the residual path, the dynamic scaling mechanism comprising an input encoding unit, a context encoding unit, and a scaling function unit;

[0023] S33, the input encoding unit performing linear mapping on x t to generate an input representation e t , the context encoding unit receiving a time window input sequence {x t -k,...,x t +k} and generating a context representation c t through a bidirectional gated recurrent unit, wherein k is the left and right expansion step number of the time window;

[0024] S34, splicing e t and c t to input the scaling function unit, calculating a scaling factor ∝ t , and performing scaling operation on r t :

[0025] r' t =σ(W1e t +W2c t +b)⊙r t ;

[0026] wherein e t is the input representation, c t is the context representation, W1 and W2 are scaling weight matrices, b is a bias vector, σ is a Sigmoid function, r t is the residual path output, r' t is the scaled residual path output, and ⊙ represents element-wise multiplication.

[0027] S35, adding r' t and the main path output y t to generate the output z of the reversible residual structure at the time step tt .

[0028] Optionally, the S4 specifically comprises:

[0029] S41, setting a time gating mechanism in the reversible residual structure in each modal subnetwork, the time gating mechanism being composed of a time embedding unit, a gating coefficient generation unit and a residual adjustment unit;

[0030] S42, the time embedding unit receives a timestamp t of a time step t, and generates a time representation τ of a dimension d:

[0031]

[0032] wherein i is a dimension index of the time representation, d is an encoding dimension, t is a time step index, is an i-th dimension time encoding value;

[0033] S43, the gating coefficient generation unit receives τ t and an output representation h t of a previous time step on a residual path, t inputs to a gating function to generate a gating coefficient γ

[0034] γ t =σ(W τ τ t +W h h t +b);

[0035] wherein τ t is the time representation, h t is the previous time step residual path output, W τ and W h are mapping matrices, b is a bias vector, σ is a Sigmoid function, and γ t is a gating coefficient vector;

[0036] S44, the residual adjustment unit receives a residual path output r t of the time step t, performs element-wise multiplication of γ t and r t , and outputs a gating adjustment result r″ t ;

[0037] S45, taking r″ t as the final output of the residual path of the time step t, and inputting it to the reversible residual structure for generating a structure output of the time step t.

[0038] Optionally, the S5 specifically comprises:

[0039] S51, adjust the residual of each modal sub-network output and record the adjusted modal features as temperature modal features, pressure modal features, speed modal features, current modal features and control signal modal features respectively;

[0040] S52, input each modal feature into the encoding transformation module of the reversible mapping structure, extract the interaction relationship between modalities and generate a fusion feature representation;

[0041] S53, record the intermediate mapping state of each time step when performing encoding transformation, which is used for reverse reconstruction process call;

[0042] S54, output the fusion feature representation as an output and input it into the forward transformation module.

[0043] Optionally, the S6 specifically includes:

[0044] S61, input the fusion feature representation sequence output in step S5 into the forward transformation module, the forward transformation module is composed of stacked affine coupling units and reversible normalization units, perform forward mapping operation on each time step, output an intermediate representation sequence, and record the mapping state at each layer;

[0045] S62, input the intermediate representation sequence and the recorded mapping state into the reverse reconstruction module, the reverse reconstruction module structure is symmetrical to the forward transformation module, and the approximate input sequence is restored;

[0046] S63, for each time step t, the input is x t , and the reconstruction output is Calculate the residual value e t :

[0047]

[0048] Where F t is the fusion feature representation corresponding to the time step t, is the forward mapping function, is the reverse reconstruction function based on the recorded state, x t is the original input of the time step t, W r is the residual weight matrix, b r is the residual bias vector, e t is the residual scalar value;

[0049] S64, construct a residual sequence from the residual value sequence and input it into the risk score module to perform scoring processing.

[0050] Optionally, the S7 specifically includes:

[0051] S71, construct a risk score module, the risk score module includes a sliding window extraction unit, a residual aggregation unit and a risk judgment unit;

[0052] S72, set the residual sequence as {e1, e2,..., e T}, the sliding window extraction unit scans the residual sequence with a sliding window of length w and step s, and extracts the residual subsequence starting at time step t;

[0053] S73, the residual aggregation unit performs exponential decay weighting on the residual values in each window, and calculates the original score r t :

[0054]

[0055] where e t+i represents the residual value at time step t+i, a is the initial weight, β is the decay factor, r t is the residual weighted score at time step t;

[0056] S74, input r t into a nonlinear normalization function to calculate the normalized score R t :

[0057]

[0058] where W r is a scaling parameter, γ is an amplification coefficient, b is a bias constant, tanh is a hyperbolic tangent function, exp is an exponential function, and R t is the normalized risk score at time step t;

[0059] S75, the risk judgment unit compares R t with a threshold θ, and outputs a high-risk label when R t is greater than or equal to θ, and outputs a low-risk label when R t is less than θ.

[0060] Optionally, the S8 specifically includes:

[0061] S81, set the attribution module, receive the residual value sequence and the original input sequence, and establish the response relationship between the input features and the residual values;

[0062] S82, set the original input sequence as X = [x1, x2,..., x T ], and the reconstruction sequence as the residual sequence is E = [e1, e2,..., e T ], and the input feature dimension contribution vector is defined as Calculate the jth dimension contribution value:

[0063]

[0064] where x t,jan input feature jth dimension component representing a time step t, a partial derivative of a residual value to a component, c j an attribution score of an input feature jth dimension;

[0065] S83, performing normalization processing on each component in the contribution vector C to generate a key factor weight vector C', and sorting the key factor weight vector C' in descending order of weight, and outputting a key factor sorting result.

[0066] The present application has the following beneficial effects:

[0067] The deep learning-based thermal power plant operation risk intelligent early warning method can effectively make up for the deficiencies of the prior art in multi-modal feature modeling, feature fidelity and risk interpretability. By introducing a reversible residual structure, each modal input feature can maintain reversible mapping of information during feature extraction, avoiding the common feature loss problem in deep neural networks. The dynamic residual scaling mechanism can adaptively adjust the residual amplitude according to the input features of each modal, making the model have stronger adaptability to data of different dimensions and scales. The time gating mechanism further introduces dynamic control ability in the time dimension, so that the model can finely adjust the sensitivity to changes at different time steps, enhancing the modeling ability for non-stationary operation processes.

[0068] In addition, by recording the intermediate state in the residual reconstruction process and combining the sliding time window to aggregate the residual value, a stable and continuous risk discrimination mechanism can be established, and the robustness of the risk score result is enhanced. The finally set fault key factor inversion module can construct the contribution score of the input feature dimension based on the residual derivative information, realize key factor sorting, and help thermal power plant operators quickly locate the source of potential risks. The overall method realizes the whole process modeling from data acquisition, feature extraction, risk scoring to abnormal attribution, and improves the intelligent level and operability of thermal power plant operation risk identification. BRIEF DESCRIPTION OF DRAWINGS

[0069] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application, and do not constitute a limitation on the application. In the drawings:

[0070] Fig. 1 a flowchart of a deep learning-based thermal power plant operation risk intelligent early warning method according to the present application;

[0071] Fig. 2 a modal subnetwork and a reversible residual structure diagram of a deep learning-based thermal power plant operation risk intelligent early warning method according to the present application;

[0072] Fig. 3A fault key factor inversion module structure diagram of a deep learning-based power plant operation risk intelligent early warning method is proposed for the present application. DETAILED DESCRIPTION

[0073] The present application will now be further described in greater detail in connection with the accompanying drawings. These drawings are all simplified schematic views which only show the basic structure of the present application in a schematic manner, and thus only show the components relevant to the present application.

[0074] REFERENCE Figs. 1-3 A deep learning-based power plant operation risk intelligent early warning method includes the following steps:

[0075] S1, collect operation data in the operation process of the power plant, and construct a multi-modal time sequence input sequence;

[0076] S2, standardize the multi-modal time sequence input sequence, and input each kind of operation data into the corresponding modal subnetwork;

[0077] S3, set a reversible residual structure in each modal subnetwork to extract modal features, introduce a dynamic scaling mechanism in the residual path, calculate a scaling factor according to the input data, and adjust the residual output amplitude;

[0078] S4, introduce a time gating mechanism in the reversible residual structure, calculate a gating coefficient according to the time stamp, and adjust the residual response of each time step;

[0079] S5, fuse the modal features adjusted by the residual through a reversible mapping structure to form a fused feature representation;

[0080] S6, perform a forward transformation operation on the fused feature representation, record the intermediate state, and perform reverse reconstruction combined with the intermediate state, calculate the residual value between the original input sequence and the reconstructed sequence;

[0081] S7, construct a risk score module, aggregate the residuals using a sliding time window to generate a score sequence, compare it with a preset threshold, and output a risk level label;

[0082] S8, set an attribution module in the reverse reconstruction path, analyze the input features based on the residual value, and output a key factor ranking result.

[0083] The present application constructs a deep learning early warning process driven by multi-modal data, integrates a reversible residual structure, a time gating mechanism and a reconstruction residual calculation link, and forms an end-to-end risk scoring and factor attribution system.

[0084] In the present embodiment, the operation data includes temperature data, pressure data, speed data, current data and control signal data.

[0085] The application identifies five key types of thermal power plant operation data, forms a comprehensive modeling basis for complex working conditions, and ensures the pertinence and integrity of the subsequent modal analysis structure.

[0086] In the embodiment, the S2 specifically includes:

[0087] S21, each operation data in the multi-modal time sequence input sequence is respectively standardized;

[0088] S22, the standardized temperature data is input into the temperature modal subnetwork, the pressure data is input into the pressure modal subnetwork, the speed data is input into the speed modal subnetwork, the current data is input into the current modal subnetwork, and the control signal data is input into the control signal modal subnetwork.

[0089] The application provides a clear data shunting path for subsequent structure processing by inputting different types of operation data into corresponding modal subnetworks after standardization processing, and enhances the perception and resolution of the model for physical quantities.

[0090] In the embodiment, the S3 specifically includes:

[0091] S31, a reversible residual structure is arranged in each modal subnetwork, the reversible residual structure includes a main path and a residual path, the residual path receives the modal input x t after the standardization processing in step S2 t , and generates a residual output r t through a neural network structure sequentially composed of a linear transformation layer, an activation function layer and a normalization layer; t ;

[0092] S32, a dynamic scaling mechanism is embedded in the residual path, the dynamic scaling mechanism is composed of an input encoding unit, a context encoding unit and a scaling function unit;

[0093] S33, the input encoding unit performs linear mapping to generate an input representation e t , the context encoding unit receives a time window input sequence {x t -k,...,x t +k} and generates a context representation c t through a bidirectional gated recurrent unit, wherein k is the left and right expansion step number of the time window;

[0094] S34, after splicing e t and c t , a scaling factor a t is calculated, and a scaling operation is performed on r t :

[0095] r' t = σ(W1e t +W2ct +b)⊙r t ;

[0096] wherein e t is an input representation, c t is a context representation, W1 and W2 are scaling weight matrices, b is a bias vector, σ is a Sigmoid function, r t is a residual path output, r' t is a scaled residual path output, and ⊙ denotes element-wise multiplication.

[0097] S35, adds r' t to the main path output y t to generate the output z t of the invertible residual structure at time step t.

[0098] The application constructs a dynamic residual scaling mechanism containing a context encoding and scaling function in the modal subnetwork, realizes fine-grained adjustment of modal residual information, and effectively improves the response accuracy of the model to abnormal features.

[0099] In the embodiment, the S4 specifically comprises:

[0100] S41, a time gating mechanism is arranged in the invertible residual structure in each modal subnetwork, and the time gating mechanism is composed of a time embedding unit, a gating coefficient generation unit and a residual adjustment unit;

[0101] S42, the time embedding unit receives the timestamp t at time step t to generate a time representation τ of dimension d:

[0102]

[0103] wherein i is the dimension index of the time representation, d is the encoding dimension, t is the time step index, is the i-th dimension time encoding value;

[0104] S43, the gating coefficient generation unit receives τ t and the output representation h t at the previous time step of the residual path, and inputs them into a gating function to generate a gating coefficient γ t :

[0105] γ t = σ(W τ τ t +W h h t +b);

[0106] wherein τ t is the time representation, h t is the previous time step residual path output, W τ and Wh is a mapping matrix, b is a bias vector, and t is a gating coefficient vector.

[0107] S44, the residual adjustment unit receives the residual path output r of the time step t t , and t and r t performs element-wise multiplication to output the gating adjustment result r t .

[0108] S45, r t is taken as the final output of the residual path of the time step t and is input to the reversible residual structure for generating the structural output of the time step t.

[0109] The application realizes the residual path regulation mechanism at the time step granularity by fusing the time embedding and the gating function, so that the model can perceive the time sequence change characteristics of the running data and dynamically adjust the structural output.

[0110] In the embodiment, the S5 specifically includes:

[0111] S51, the residual adjusted modal features output by each modal subnetwork are respectively recorded as temperature modal features, pressure modal features, speed modal features, current modal features and control signal modal features;

[0112] S52, each modal feature is input to the encoding transformation module of the reversible mapping structure to extract the inter-modal interaction relationship and generate a fused feature representation;

[0113] S53, the intermediate mapping state of each time step is recorded when the encoding transformation is performed, which is used for calling in the reverse reconstruction process;

[0114] S54, the fused feature representation is taken as the output and is transmitted to the forward transformation module.

[0115] The application fuses each modal feature into a unified representation by using the reversible mapping structure, and records the intermediate state in the encoding stage, thereby providing structural support for subsequent reversible reconstruction and risk estimation.

[0116] In the embodiment, the S6 specifically includes:

[0117] S61, the fused feature representation sequence output in step S5 is input to the forward transformation module, the forward transformation module is composed of stacked affine coupling units and reversible normalization units, the forward mapping operation is performed on each time step, an intermediate representation sequence is output, and the mapping state is recorded at each layer.

[0118] S62, input the intermediate representation sequence and the recorded mapping state into a reverse reconstruction module, the reverse reconstruction module is symmetrical in structure to the forward transformation module, and an approximate input sequence is recovered;

[0119] S63, for each time step t, input is x t , and the reconstruction output is Calculate the residual value e t :

[0120]

[0121] Wherein, F t is the fusion feature representation corresponding to the time step t, is a forward mapping function, is a reverse reconstruction function executed based on the recorded state, x t is the original input of the time step t, W r is a residual weight matrix, and b r is a residual bias vector, e t is a residual scalar value;

[0122] S64, the residual value sequence is formed into a residual sequence, and a risk score module is input to execute scoring processing.

[0123] The application adopts a symmetrical structure to complete feature reconstruction, and constructs a scoring process of fusion residual value, so that the model can quantify the risk based on the structure recovery error, and form a continuous residual score sequence.

[0124] In the embodiment, the S7 specifically comprises:

[0125] S71, a risk score module is constructed, the risk score module comprises a sliding window extraction unit, a residual aggregation unit and a risk judgment unit;

[0126] S72, the residual sequence is set as {e1, e2,..., e T}, the sliding window extraction unit scans the residual sequence by using a sliding window with a length of w and a step of s, and extracts a residual subsequence starting from the time step t;

[0127] S73, the residual aggregation unit performs exponential decay weighting on the residual values in each window, and calculates the original score r t :

[0128]

[0129] Wherein, e t+i represents the residual value of the time step t+i, α is an initial weight, β is a decay factor, r t is the residual weighted score of the time step t;

[0130] S74, rt Input a nonlinear normalization function to calculate the normalized score R t :

[0131]

[0132] where W r is a scaling parameter, gamma is an amplification coefficient, b is a bias constant, tanh is a hyperbolic tangent function, exp is an exponential function, R t is the normalized risk score at time step t, and

[0133] S75, the risk judgment unit compares R t with a threshold value theta, R t greater than or equal to theta outputs a high-risk label, and R t less than theta outputs a low-risk label.

[0134] The present application introduces a sliding window and an exponential weighting mechanism on the basis of residual scores, and combines nonlinear mapping and threshold judgment logic to realize risk level label output under multiple scales.

[0135] In the present embodiment, the S8 specifically comprises:

[0136] S81, an attribution module is set up to receive residual value sequences and original input sequences, and to establish a response relationship between input features and residual values;

[0137] S82, the original input sequence is set as X = [x1, x2,..., x T ], and the reconstruction sequence is The residual sequence is E = [e1, e2,..., e T ], and the input feature dimension contribution vector C is defined The jth dimension contribution value is calculated:

[0138]

[0139] where x t,j represents the jth dimension component of the input feature at time step t, represents the partial derivative of the residual value with respect to the component, and c j represents the attribution score of the jth dimension of the input feature.

[0140] S83, normalization processing is performed on each component in the contribution vector C to generate a key factor weight vector C', and the key factor weight vector C' is sorted in descending order of weight, and the key factor sorting result is output.

[0141] The present application constructs a key factor contribution vector based on the gradient response between input residuals, and quantifies the influence weight of the key factor by normalization and sorting method to output an ordered risk factor set.

[0142] Example 1

[0143] To verify the feasibility of the application in implementation, the application is applied to a typical coal-fired power plant in East China with a installed capacity of 2x600MW, which is in a high load operation state for a long time, and is equipped with SCADA and DCS control systems, and has complete multi-source operation data acquisition capability. The plant feedbacks that the existing alarm system based on static threshold has problems such as response delay and key factor positioning. Especially during the night shift period, the abnormal judgment ability of the on-duty personnel is limited, resulting in frequent false positives and false negatives.

[0144] In the implementation process of the application, the system collects five types of key operation data: temperature (T), pressure (P), speed (S), current (I) and control signal (C) through industrial protocol. The data sampling frequency is 1 time / minute, and a 60-minute sliding window is used to construct the time series input. The input multi-modal deep network has a model structure including five modal sub-networks, each of which integrates a reversible residual structure, and introduces a dynamic residual scaling mechanism and a time gating mechanism in the residual path. Each type of modal feature is input into the reversible mapping structure after independent extraction, fused, compressed by forward transformation, and the intermediate state is recorded. The residual matrix of the predicted sequence and the real input sequence is obtained by backward reconstruction. The system uses a sliding time window to aggregate the residual to generate a risk score sequence, and judges the risk level according to the score and the set threshold. At the same time, the key influencing factor sorting result is output through the attribution module.

[0145] In the process of system deployment, the data of multiple key events are recorded. For example, on April 3, 2024, the system detected an anomaly in the main cylinder area, with a risk score peak of 0.83, and the model successfully issued a warning 26 minutes before the anomaly occurred. The risk score of this event increased significantly, and the model timely identified the problem, and according to the different input features, the corresponding factor sorting was output, helping the operator to quickly locate the abnormal source. Specifically, the main cylinder pressure P3, the main feed water current I5 and the reheat temperature T8 become the most important influencing factors, accounting for 38%, 27% and 18% respectively. In this event, the accurate warning of the system provides timely response time for the operator, successfully avoiding the shutdown loss caused by abnormal development.

[0146] Table 1 is the key event data collected during the actual deployment of the system, which reflects the effectiveness of the model warning of the application. The data in the table reflects the specific abnormal events detected by the system, and shows the risk score peak, the warning advance time and the actual impact duration:

[0147] Table 1 Abnormal event and model response data

[0148]

[0149] The data in the above table shows 6 typical abnormal events and the response effect of the system. In the main cylinder abnormal event on April 3, 2024, the system issued a warning 26 minutes before the abnormality occurred, with a risk score peak of 0.83, indicating that the event was of high severity. At this point in time, the model successfully identified "main cylinder pressure P3" as the main factor and further optimized the warning effect through dynamic adjustment of the response. The actual impact duration was 90 minutes, indicating that although the event lasted for a long time, the timely warning enabled the plant to take effective measures to avoid serious losses.

[0150] In addition, other events in the table also reflect the warning capability of the method of the present application under different working conditions. For example, the superheater over-temperature event on April 22, 2024, the system issued a warning 31 minutes in advance, timely capturing the temperature change of the superheater, avoiding equipment failure caused by overheating, and through the key factor attribution analysis of the system, the plant was able to accurately locate the root cause of the temperature change, greatly reducing the impact of the failure.

[0151] After three months of continuous implementation, the warning method of the present application recorded 12 high-risk events with risk scores exceeding the threshold in actual application, of which 9 real abnormalities were successfully identified. The average warning time was 21.4 minutes, and in all key events, the system effectively improved the response time and warning accuracy, compared with the original risk warning system based on static threshold, the response time was improved by about 60%, the false positive rate was reduced by about 37%, significantly improving the safety and stability of operation.

[0152] The implementation of the present application verifies the feasibility and effectiveness of applying an intelligent warning method based on deep learning in the operation environment of a thermal power plant. Through deep modeling and accurate warning of multi-modal data, the present application provides strong technical support for the intelligent management of thermal power plants, further improving the safety, stability and management efficiency of thermal power plants.

[0153] The above description is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A deep learning-based intelligent early warning method for operation risk of a thermal power plant, characterized in that, The method comprises the following steps: S1, collecting operation data in the operation process of a thermal power plant to construct a multi-modal time sequence input sequence; S2, performing standardization processing on the multi-modal time sequence input sequence, and inputting each operation data into a corresponding modal subnetwork; S3, setting a reversible residual structure in each modal subnetwork to extract modal features, and introducing a dynamic scaling mechanism in a residual path to calculate a scaling factor according to input data and adjust a residual output amplitude; S4, introducing a time gating mechanism in the reversible residual structure to calculate a gating coefficient according to a time stamp and adjust a residual response of each time step; S5, fusing the modal features adjusted by the residual to form a fused feature representation through a reversible mapping structure; S6, performing a forward transformation operation on the fused feature representation, recording an intermediate state, and combining the intermediate state to perform reverse reconstruction to calculate a residual value between an original input sequence and a reconstructed sequence; S7, constructing a risk score module, aggregating residuals by using a sliding time window to generate a score sequence, comparing the score sequence with a preset threshold, and outputting a risk level label; S8, setting an attribution module in the reverse reconstruction path, analyzing input features based on the residual value, and outputting a key factor ranking result.

2. The deep learning-based intelligent early warning method for operation risk of a thermal power plant according to claim 1, characterized in that, The operation data includes temperature data, pressure data, speed data, current data and control signal data.

3. The deep learning-based intelligent early warning method for operation risk of a thermal power plant according to claim 1, characterized in that, The S2 specifically comprises: S21, performing standardization processing on each operation data in the multi-modal time sequence input sequence respectively; S22, inputting the standardized temperature data into a temperature modal subnetwork, inputting the pressure data into a pressure modal subnetwork, inputting the speed data into a speed modal subnetwork, inputting the current data into a current modal subnetwork, and inputting the control signal data into a control signal modal subnetwork.

4. The deep learning-based intelligent early warning method for operation risk of a thermal power plant according to claim 1, characterized in that, The S3 specifically comprises: S31, setting a reversible residual structure in each modal subnetwork, the reversible residual structure comprising a main path and a residual path, the residual path receiving the modal input x after the standardization processing in step S2 t ; generating the residual output r through a neural network structure sequentially composed of a linear transformation layer, an activation function layer, and a normalization layer t . S32, embedding a dynamic scaling mechanism in the residual path, and the dynamic scaling mechanism comprises an input encoding unit, a context encoding unit and a scaling function unit; S33, the input encoding unit encodes x t performing linear mapping to generate input representation e t The context encoding unit receives the time window input sequence {x t -k,...,x t +k} and generates context representation c t where k is the left and right expansion step number of the time window; S34, e t With c t The input scaling function unit after splicing, calculate scaling factor ∝ t And r t Perform scaling operation: r' t = σ(W1e t +W2c t +b)⊙r t ; where e t is the input representation, c t is the context representation, W1 and W2 are scaling weight matrices, b is a bias vector, σ is a Sigmoid function, r t is the residual path output, r' t is the scaled residual path output, and ⊙ denotes element-wise multiplication. S35, r' = r + y t t ​​ 5. The deep learning-based intelligent early warning method for operation risk of a thermal power plant according to claim 1, characterized in that, The S4 specifically comprises: S41, setting a time gating mechanism in the reversible residual structure in each modal subnetwork, and the time gating mechanism comprises a time embedding unit, a gating coefficient generation unit and a residual adjustment unit; S42, the time embedding unit receives a time stamp t of a time step t to generate a time representation τ with a dimension d: where i is the dimension index of the time representation, d is the encoding dimension, t is the time step index, is the i-th dimensional time encoding value; S43, the gating coefficient generation unit receives τ t with the output representation h t at the previous time step on the residual path t : gamma t = sigma(W τ tau t + W h h t + b); Where, τ f For time representation, h t W is the output of the residual path from the previous time step. τ With W h Let be the mapping matrix, b be the bias vector, σ be the sigmoid function, and γ be the bias vector. f This is the gating coefficient vector; S44, the residual adjustment unit receives the residual path output r of the time step t t , γ t , and r t , performs element-wise multiplication, and outputs the gating adjustment result r" t ; S45, r” = r' - r t As a residual path final output for time step t, it is input to the invertible residual structure for generating a structure output for time step t.

6. The deep learning-based intelligent early warning method for operation risk of a thermal power plant according to claim 1, characterized in that, The S5 specifically comprises: S51, recording the residual adjusted modal features output by each modal subnetwork as temperature modal features, pressure modal features, speed modal features, current modal features and control signal modal features respectively; S52, inputting each modal feature into an encoding transformation module of the reversible mapping structure to extract inter-modal interaction relationships and generate a fused feature representation; S53, recording an intermediate mapping state of each time step when performing the encoding transformation, which is used for calling in a reverse reconstruction process; S54, taking the fused feature representation as an output and transmitting it into a forward transformation module.

7. The deep learning-based intelligent early warning method for operation risk of a thermal power plant according to claim 1, characterized in that, The S6 specifically comprises: S61, inputting the fused feature representation sequence output in step S5 into a forward transformation module, and the forward transformation module is composed of stacked affine coupling units and reversible normalization units, performs a forward mapping operation on each time step, outputs an intermediate representation sequence, and records a mapping state at each layer; S62, input the intermediate representation sequence and the recorded mapping state into a reverse reconstruction module, the reverse reconstruction module is symmetrical to the forward transformation module, and an approximate input sequence is recovered; S63, for each time step t, the input is x t , the reconstructed output is The residual value e t is calculated: where F t is the fused feature representation for time step t, is the forward mapping function, is the inverse reconstruction function executed based on the recording state, x t is the original input for time step t, W r is the residual weighting matrix, b r is the residual bias vector, e t is the residual scalar value; S64, the residual value sequence is constructed into a residual sequence, and a risk score module is input to perform a scoring process.

8. The deep learning-based intelligent early warning method for operation risk of a thermal power plant according to claim 1, characterized in that, The S7 specifically includes: S71, constructing a risk score module, the risk score module includes a sliding window extraction unit, a residual aggregation unit, and a risk judgment unit; S72, set the residual sequence as {e1, e2,..., e T}, the sliding window extraction unit scans the residual sequence by using a sliding window with a length of w and a step of s, and extracts a residual sub-sequence starting at time step t; S73, the residual aggregation unit performs exponential decay weighting on the residual values within each window, and calculates the original score r t : where e t+i denotes the residual value at time step t+i, a is an initial weight, β is a decay factor, r t is the residual weighted score at time step t. S74, r t inputting a non-linear normalization function, calculating a normalized score R t : where W r is a scaling parameter, γ is a magnification factor, b is a bias constant, tanh is the hyperbolic tangent function, exp is the exponential function, R t is the normalized risk score at time step t; S75, the risk judgment unit compares R t with the threshold value θ, R t outputs a high-risk label when R t outputs a low-risk label when R 9. The deep learning-based intelligent early warning method for operation risk of a thermal power plant according to claim 1, characterized in that, The S8 specifically includes: S81, setting an attribution module, receiving a residual value sequence and an original input sequence, and establishing a response relationship between input features and residual values; S82, set the original input sequence as X = [x1, x2,..., x T ], the reconstruction sequence as the residual sequence as E = [e1, e2,..., e T ], define the input feature dimension contribution vector calculate the jthdimension contribution value: wherein x t,j denotes the input feature j-th dimension component at time step t, denotes the partial derivative of the residual value with respect to the component, c j denotes the attribution score of the input feature j-th dimension; S83, performing normalization processing on each component in the contribution vector C to generate a key factor weight vector C', and sorting the key factor weight vector C' from large to small according to the weight, and outputting a key factor sorting result.

Images