Intelligent start-stop control method and system for a concentrated solar power system based on a distributed control system

Abstract

The application provides a DCS-based intelligent start-stop control method and system for a photo-thermal power generation system, relates to the technical field of photo-thermal power generation, and comprises the following steps: collecting real-time operation data through a DCS, extracting features through data standardization and wavelet transformation, extracting time sequence features by using a variational recursive neural network, generating a reconstruction sequence by using an auto-encoder to screen key time segments, extracting decision rules by using knowledge distillation, constructing a strategy network to generate a control strategy, and deploying the control strategy to a DCS control unit. The application can improve the start-stop control precision of the photo-thermal power generation system, reduce the operation risk of the system, and prolong the service life of equipment.

Description

Technical Field

[0001] This invention relates to the field of concentrated solar power (CSP) technology, and in particular to a method and system for intelligent start-stop control of CSP systems based on DCS. Background Technology

[0002] Concentrated solar power (CSP), as a clean and renewable energy technology, converts solar energy into electricity by focusing solar energy. It is environmentally friendly and sustainable, and has become an important option for global energy transition. CSP systems are complex, comprising multiple subsystems including solar collectors, thermal storage systems, heat exchange systems, and power generation systems. Their operation is affected by various factors such as weather conditions, equipment performance, and grid demand. The system's start-up and shutdown processes have a significant impact on equipment lifespan and operating efficiency; reasonable start-up and shutdown control strategies can extend equipment lifespan, improve system operating efficiency, and reduce operating costs.

[0003] Traditional concentrated solar power (CSP) system start-up and shutdown control relies primarily on manual experience and fixed rules. Operators manually execute start-up and shutdown operations based on weather forecasts, system operating status, and grid dispatch information. While the application of distributed control systems (DCS) has improved the automation level of CSP systems, current control methods still have several shortcomings: Existing CSP system start-up and shutdown control lacks the ability to deeply mine complex time-series data, failing to effectively extract hidden patterns and correlation features, resulting in insufficient system adaptability to environmental changes and equipment status, and a tendency to make decision-making errors under complex operating conditions. Traditional control methods struggle to effectively transfer and accumulate knowledge and experience; system operating experience often exists as tacit knowledge in the minds of operators, lacking a systematic extraction and application mechanism. This makes it difficult to fully utilize valuable historical optimization experience in new systems, limiting the continuous optimization of control strategies. Current start-up and shutdown control strategies generally lack adaptive learning capabilities, failing to automatically adjust control parameters and decision logic according to dynamic changes in system operating status. When facing new operating conditions or equipment status changes, the control effect is often unsatisfactory, making it difficult to achieve continuous optimization of system performance. Summary of the Invention

[0004] The present invention provides a method and system for intelligent start-stop control of a concentrated solar power generation system based on DCS, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides an intelligent start-stop control method for a concentrated solar power (CSP) system based on a DCS, comprising: Real-time operating data of the solar thermal power generation system is collected by a distributed control system (DCS); the real-time operating data is standardized; wavelet transform is used to decompose the standardized data into multi-scale components at different frequency scales; principal component analysis is used to reduce the dimensionality of the feature components to obtain the principal feature vector; and the principal feature vector is combined with the real-time operating data to form the original feature sequence. The temporal features of the original feature sequence are extracted using a variational recurrent neural network, and the temporal features are input into an autoencoder to generate a feature reconstruction sequence; the reconstruction error between the original feature sequence and the reconstruction sequence is calculated, and key time segments are selected based on the reconstruction error; the key time segments are input into a recurrent neural network to generate a dynamic feature representation of the system state; The dynamic feature representation is input into a knowledge distillation network to extract start-stop decision rules; a policy network for generating start-stop control strategies is constructed, the policy network adopts a deep neural network structure, and the start-stop decision rules are encoded as the initial parameters of the policy network; a contrastive learning method is used to calculate the similarity score between the current decision and historical experience, and the learning direction of the policy network is adjusted based on the similarity score; The adjusted strategy network is deployed to the field control unit of the DCS. The field control unit sends start-stop control commands to the solar thermal power generation system according to the strategy output by the strategy network, thereby realizing intelligent start-stop control of the system.

[0006] Wavelet transform is used to perform multi-scale decomposition on the standardized data to obtain feature components at different frequency scales, including: Local statistical features are generated based on the standardized data calculated using a sliding time window. These local statistical features are then fed into a trained neural network to obtain the weight coefficients for each basis function. The weight coefficients are then multiplied by each basis function in a preset set of basis functions and summed to obtain the dynamic wavelet basis function. The standardized data is subjected to one-dimensional convolution operation through the dynamic wavelet basis function to generate multi-layer decomposition coefficients. For each decomposition coefficient, the information entropy of its approximate coefficient is calculated using the entropy function, and the energy ratio of its detail coefficient to the total energy is calculated to obtain the energy ratio. The information entropy and the energy ratio are multiplied by the corresponding preset weights and then added to obtain the evaluation index of the layer. The layer with the largest evaluation index is determined as the optimal decomposition layer. The standardized data is subjected to discrete wavelet transform through the dynamic wavelet basis function. The data is decomposed layer by layer according to the optimal decomposition level to obtain the approximation coefficients and detail coefficients of each level. The weighted average of the detail coefficients of each level and the detail coefficients of its adjacent upper and lower levels is calculated. The detail coefficients of the level are subtracted from the weighted average and the sum of squares is obtained to obtain the scale constraint value. The value of the detail coefficients is adjusted according to the scale constraint value until the scale constraint value is less than the preset scale threshold. The adjusted detail coefficients are then subjected to one-dimensional convolution operations using the dynamic wavelet basis function to reconstruct the feature components at different frequency scales.

[0007] The temporal features of the original feature sequence are extracted using a variational recurrent neural network, and the temporal features are input into an autoencoder to generate a feature reconstruction sequence, including: Based on the feature vector of the current time step and the hidden state of the previous time step in the original feature sequence, the mean parameter and variance parameter of the variational posterior distribution are calculated. The variational posterior distribution is determined using the mean parameter and the variance parameter. The hidden variable of the current time step is sampled from the variational posterior distribution. The feature vector at the current time, the hidden state at the previous time, and the hidden variable at the current time are input into the variational recursive unit to obtain the hidden state at the current time. The variational posterior distribution calculation and the hidden state calculation are repeated until all feature vectors in the original feature sequence are processed to obtain the hidden state sequence corresponding to each time. The similarity between any two hidden states in the hidden state sequence is calculated to obtain a similarity score. The similarity score is normalized to obtain an attention weight sequence. The attention weight sequence is weighted and summed with the hidden state sequence at the corresponding time to obtain a context vector sequence. The hidden state sequence and the context vector sequence are concatenated along the temporal dimension and then input into the encoder to obtain initial temporal features. One-dimensional convolution operations with different kernel sizes are performed on the initial temporal features to obtain multiple scale features. The multiple scale features are weighted and summed with preset scale weights to obtain fused features. The fused features are input into the decoder to generate a feature reconstruction sequence.

[0008] Calculating the reconstruction error between the original feature sequence and the reconstructed sequence, and selecting key time segments based on the reconstruction error, includes: The Euclidean distance between the original feature sequence and the reconstructed sequence at each time step is calculated to obtain the reconstruction error sequence. A time window of a preset size is set at each time step, and the reconstruction error values ​​within the time window are accumulated to obtain the local error value at that time step, thus forming a local error sequence. Each time step in the local error sequence is treated as a node in the graph structure. The feature vector corresponding to each node at that time step is extracted. The cosine similarity between the feature vectors of any two adjacent nodes is calculated to obtain the feature correlation coefficient. The feature correlation coefficient is then filled into the corresponding position of the initial edge weight matrix. The difference between the timestamps of any two adjacent nodes is calculated. The difference is then substituted into the Gaussian kernel function to obtain the time distance decay factor. The time distance decay factor is then filled into the corresponding position of the time weight matrix. The initial edge weight matrix is ​​multiplied by the corresponding element in the time weight matrix to obtain the final edge weight matrix. According to the edge weight matrix, each node obtains a local error value from its neighboring nodes, multiplies the local error value with the corresponding edge weight and sums them to obtain the update error value of the node, and calculates the update error values ​​of all nodes in turn to form an update error sequence. The local error sequence and the error values ​​at corresponding times in the updated error sequence are weighted and combined to obtain the cumulative error score sequence. The time segments in the cumulative error score sequence that are less than a preset score threshold are merged to obtain the final key time segments.

[0009] The dynamic feature representation is input into the knowledge distillation network. Extract start / stop decision rules, including: Calculate the query matrix, key matrix, and value matrix of the dynamic feature representation, multiply the query matrix by the transpose of the key matrix and divide by the square root of the feature dimension, input the calculation result into the normalized exponential function to obtain the attention weight, and multiply the attention weight by the value matrix to obtain the attention feature; Multiple temporal convolutional layers of different scales are constructed for the dynamic feature representation. Each temporal convolutional layer extracts feature patterns within different time windows to obtain convolutional features. The attention features and the convolutional features are fused with adaptive weights to obtain enhanced features. A dual-branch structure of teacher network and student network is constructed, and start / stop decision probability distributions corresponding to teacher network and student network are generated based on the enhanced features. Samples with the same start / stop label are constructed as positive sample pairs, and samples with different start / stop labels are constructed as negative sample pairs. The contrast loss between the positive sample pairs and the negative sample pairs is calculated. The KL divergence between the start / stop decision probability distribution of the teacher network and the start / stop decision probability distribution of the student network is used as the knowledge distillation loss. The knowledge distillation loss and the contrastive loss are weighted and combined to obtain the total loss function. Based on the total loss function, the gradient of each start-stop related feature in the student network with respect to the start-stop decision result is calculated to obtain the feature importance. The start / stop decision rules are scored based on a weighted combination of the importance of the features and the attention features, and start / stop decision rules with an importance score greater than a preset score threshold are extracted.

[0010] The similarity score between the current decision and historical experience is calculated using a contrastive learning method. The learning direction of the policy network is then adjusted based on this similarity score, including: Using a sliding time window, temporal feature sequences are extracted sequentially from the historical experience; the temporal feature sequences are input into the forward and backward networks of a bidirectional long short-term memory network to generate forward hidden state sequences and backward hidden state sequences, respectively; the forward hidden state sequences and the backward hidden state sequences are concatenated along the feature dimension to obtain a temporal representation vector. For the time-series representation vector, time-series pooling operations are performed at different time scales to obtain multi-scale feature vectors. The similarity between different feature vectors is calculated. Feature vector pairs with similarity greater than a preset positive sample threshold are stored as positive sample pairs in the positive sample set, and feature vector pairs with similarity less than a preset negative sample threshold are stored as negative sample pairs in the negative sample set; wherein, the preset positive sample threshold is greater than the preset negative sample threshold. The multi-scale feature vectors are projected into the state space and the action space respectively. The cosine distance between the feature vectors is calculated in the state space to obtain the state similarity, and the cosine distance between the feature vectors is calculated in the action space to obtain the action similarity. The state similarity and the action similarity are concatenated and input into the attention network. The attention network outputs dynamic weight coefficients. The state similarity and the action similarity are weighted and summed according to the dynamic weight coefficients to obtain a similarity score. Based on the similarity score, the sample pair with the highest similarity is selected from the set of positive samples as reference experience, and the action strategy of the reference experience is used as the learning direction for the current decision.

[0011] A second aspect of the present invention provides an intelligent start-stop control system for a concentrated solar power (CSP) system based on a DCS, comprising: The first unit is used to collect real-time operating data of the solar thermal power generation system through the distributed control system (DCS); to standardize the real-time operating data; to decompose the standardized data into multiple scales using wavelet transform to obtain feature components at different frequency scales; to reduce the dimensionality of the feature components using principal component analysis to obtain the principal feature vector; and to combine the principal feature vector with the real-time operating data to form the original feature sequence. The second unit is used to extract the temporal features of the original feature sequence using a variational recurrent neural network, input the temporal features into an autoencoder to generate a feature reconstruction sequence; calculate the reconstruction error between the original feature sequence and the reconstruction sequence, select key time segments based on the reconstruction error; and input the key time segments into a recurrent neural network to generate a dynamic feature representation of the system state. The third unit is used to input the dynamic feature representation into a knowledge distillation network to extract start-stop decision rules; construct a policy network for generating start-stop control strategies, wherein the policy network adopts a deep neural network structure, and encodes the start-stop decision rules as the initial parameters of the policy network; calculate the similarity score between the current decision and historical experience using a contrastive learning method, and adjust the learning direction of the policy network based on the similarity score; The fourth unit is used to deploy the adjusted strategy network to the field control unit of the DCS. The field control unit sends start-stop control commands to the solar thermal power generation system according to the strategy output by the strategy network to realize intelligent start-stop control of the system.

[0012] A third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] Fourth aspect of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0014] The beneficial effects of this application are as follows: This invention collects real-time data through a distributed control system and performs multi-scale decomposition using wavelet transform. Then, it extracts principal feature vectors through principal component analysis to form an original feature sequence containing the key characteristics of the system. This significantly improves the feature extraction capability for complex operating conditions of solar thermal power generation systems, accurately captures the dynamic changes in the system's operating status, and provides a reliable data foundation for intelligent start-stop control.

[0015] This invention employs a variational recurrent neural network combined with an autoencoder to extract time-series features and filter key time segments. Then, a recurrent neural network is used to generate a dynamic feature representation of the system state, enabling accurate identification of key states during the start-up and shutdown of a solar thermal power generation system. This effectively solves the problem that traditional control methods struggle to cope with the nonlinear characteristics and time-varying parameters of the system, thereby improving the stability and safety of the system operation.

[0016] This invention extracts start-up and shutdown decision rules through a knowledge distillation network and dynamically adjusts the strategy network using a comparative learning method. Finally, the optimized strategy is deployed to the field control unit, realizing intelligent control of the start-up and shutdown process of the solar thermal power generation system. This significantly reduces energy consumption and equipment wear during system start-up and shutdown, extends equipment lifespan, and improves the overall operating efficiency and economic benefits of the power generation system. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent start-stop control method for a concentrated solar power (CSP) system based on DCS, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the dynamic wavelet signal analysis architecture. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the intelligent start-stop control method for a concentrated solar power (CSP) system based on a DCS, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Real-time operating data of the solar thermal power generation system is collected by a distributed control system (DCS); the real-time operating data is standardized; wavelet transform is used to decompose the standardized data into multi-scale components at different frequency scales; principal component analysis is used to reduce the dimensionality of the feature components to obtain the principal feature vector; and the principal feature vector is combined with the real-time operating data to form the original feature sequence. The temporal features of the original feature sequence are extracted using a variational recurrent neural network, and the temporal features are input into an autoencoder to generate a feature reconstruction sequence; the reconstruction error between the original feature sequence and the reconstruction sequence is calculated, and key time segments are selected based on the reconstruction error; the key time segments are input into a recurrent neural network to generate a dynamic feature representation of the system state; The dynamic feature representation is input into a knowledge distillation network to extract start-stop decision rules; a policy network for generating start-stop control strategies is constructed, the policy network adopts a deep neural network structure, and the start-stop decision rules are encoded as the initial parameters of the policy network; a contrastive learning method is used to calculate the similarity score between the current decision and historical experience, and the learning direction of the policy network is adjusted based on the similarity score; The adjusted strategy network is deployed to the field control unit of the DCS. The field control unit sends start-stop control commands to the solar thermal power generation system according to the strategy output by the strategy network, thereby realizing intelligent start-stop control of the system.

[0021] The real-time operating data in this invention can be selected from one or more data or a combination of data related to core state characteristics, key control variables, speed control variables, and operating data related to special states (special state indications).

[0022] The core state characteristics include one or more or a combination of the following: energy conversion critical state, molten salt thermal storage state, heat exchange process state, and hydrodynamic state. In one embodiment, operational data related to the energy conversion critical state includes one or more or a combination of the following: main steam pressure, main steam temperature, high-temperature reheat steam pressure, high-temperature reheat steam temperature, turbine exhaust steam well water level, main feedwater header flow rate, and main feedwater header temperature. Operational data related to the molten salt thermal storage state includes one or more or a combination of the following: cryogenic molten salt tank level, high-temperature molten salt tank level, molten salt temperature 1~N (a representative point can be selected) in the cryogenic molten salt tank, molten salt temperature 1~N (a representative point can be selected) in the high-temperature molten salt tank, molten salt temperature in the cryogenic molten salt pump outlet header, molten salt temperature in the high-temperature molten salt pump outlet header, and temperature-regulating molten salt pump outlet header temperature. Operational data related to the heat exchange process state includes one or more or a combination of the following: steam drum level, steam drum pressure, superheater outlet main steam temperature, reheater outlet steam temperature, and preheater to steam drum feedwater temperature. Operating data related to fluid dynamics include one or more of the following: main feedwater pump outlet header pressure, condensate pump outlet header pressure, cryogenic molten salt pump outlet header pressure, high-temperature molten salt pump outlet header pressure, absorber to high-temperature molten salt tank inlet pipeline pressure, and molten salt pressure returning to the cryogenic molten salt tank, or a combination thereof. Operating data related to flow balance include one or more of the following: cryogenic molten salt pump outlet molten salt flow rate, high-temperature molten salt pump outlet molten salt flow rate, deaerator inlet condensate flow rate, main steam desuperheating water flow rate, and reheat steam desuperheating water flow rate, or a combination thereof.

[0023] Key control variables include one or more data, or combinations thereof, from temperature control actuators, flow control actuators, pressure control actuators, and system switching actuators (discrete actions). In one embodiment, the operating data for the temperature control actuator includes one or more, or combinations thereof, from the main steam desuperheating water electric regulating valve control command / position, the reheat steam desuperheating water electric regulating valve control command / position, and the electric regulating valve control command / position from the temperature regulating molten salt main pipe to the high-temperature molten salt main pipe. The operating data for the flow control actuator includes one or more, or combinations thereof, from the main feedwater header electric regulating valve control command / position, the preheater inlet electric regulating valve control command / position, the cryogenic molten salt pump outlet electric regulating valve control command / position, and the high-temperature molten salt pump outlet electric regulating valve control command / position. The operating data for the pressure control actuator includes one or more, or combinations thereof, from the lower tower high-pressure differential pneumatic regulating valve control command / position and the deaerator water level pneumatic main regulating valve control command / position. The operating data in the system switching actuator (discrete action) includes one or more of the following: the opening / closing command of the bypass electric valve from the upper tower to the lower tower, the opening / closing command of the electric valve from the lower tower to the high-temperature molten salt storage tank, and the opening / closing command of the electric valve from the lower tower to the low-temperature molten salt storage tank, or a combination thereof.

[0024] In one embodiment, the speed control variables include one or more of the main feedwater pump speed, the low-temperature molten salt pump speed, the high-temperature molten salt pump speed, and the temperature-regulating molten salt pump speed, or a combination thereof.

[0025] Special status indications include one or more of the following: temperature gradient anomaly detection and liquid level safety monitoring. In one embodiment, temperature gradient anomaly detection includes steam drum inner wall temperature 1~N (temperature difference monitoring), steam drum outer wall temperature 1~N (temperature difference monitoring), temperature comparison before / after the main steam spray desuperheater, and temperature comparison before / after the reheat steam spray desuperheater. Liquid level safety monitoring includes the liquid level in the continuous blowdown expansion tank, the liquid level in the brine tank, and the liquid level in the low-load preheater.

[0026] Figure 2 This is a schematic diagram of a dynamic wavelet signal analysis architecture. In one optional implementation, wavelet transform is used to perform multi-scale decomposition on the standardized data to obtain feature components at different frequency scales, including: Local statistical features are generated based on the standardized data calculated using a sliding time window. These local statistical features are then fed into a trained neural network to obtain the weight coefficients for each basis function. The weight coefficients are then multiplied by each basis function in a preset set of basis functions and summed to obtain the dynamic wavelet basis function. The standardized data is subjected to one-dimensional convolution operation through the dynamic wavelet basis function to generate multi-layer decomposition coefficients. For each decomposition coefficient, the information entropy of its approximate coefficient is calculated using the entropy function, and the energy ratio of its detail coefficient to the total energy is calculated to obtain the energy ratio. The information entropy and the energy ratio are multiplied by the corresponding preset weights and then added to obtain the evaluation index of the layer. The layer with the largest evaluation index is determined as the optimal decomposition layer. The standardized data is subjected to discrete wavelet transform through the dynamic wavelet basis function. The data is decomposed layer by layer according to the optimal decomposition level to obtain the approximation coefficients and detail coefficients of each level. The weighted average of the detail coefficients of each level and the detail coefficients of its adjacent upper and lower levels is calculated. The detail coefficients of the level are subtracted from the weighted average and the sum of squares is obtained to obtain the scale constraint value. The value of the detail coefficients is adjusted according to the scale constraint value until the scale constraint value is less than the preset scale threshold. The adjusted detail coefficients are then subjected to one-dimensional convolution operations using the dynamic wavelet basis function to reconstruct the feature components at different frequency scales.

[0027] This invention discloses a method for multi-scale decomposition of standardized data using wavelet transform. This method is used to obtain feature components at different frequency scales, and the specific implementation is as follows.

[0028] Standardized data typically refers to the result of normalizing the original data, such as mapping the data to a range of 0 to 1, or data with a mean of 0 and a variance of 1. Assume the standardized data sequence is X=[x1, x2, ..., x...]. n ], where n is the data length.

[0029] The standardized data were processed using a sliding time window, with a window size of 16 sample points and a sliding step size of 4 sample points. Within each window, local statistical characteristics were calculated, including mean, variance, maximum, minimum, kurtosis, and skewness. For example, for the data subset [x1, x2, ..., x...] within the window... 16 The eigenvector F = [0.35, 0.12, 0.86, 0.05, 2.34, -0.15] is calculated, representing the six statistical characteristic values ​​within the window.

[0030] The aforementioned local statistical features F are input into a pre-trained neural network. This neural network consists of an input layer (6 neurons), a hidden layer (12 neurons), and an output layer (8 neurons). The network uses the ReLU activation function and is trained using backpropagation and gradient descent. The 8 neurons in the output layer correspond to 8 weight coefficients [w1, w2, ..., w8]. For example, after inputting the aforementioned feature vector F, the neural network outputs weight coefficients W=[0.23,0.18, 0.15, 0.12, 0.10, 0.09, 0.08, 0.05].

[0031] The pre-defined basis function set includes eight commonly used wavelet basis functions, such as Haar, Daubechies4, Symlet4, and Coiflet3. The dynamic wavelet basis function is obtained by multiplying the weight coefficients of the neural network output by each basis function and then summing the results. For example, if the value of the first basis function at a certain point is 0.5, then the dynamic wavelet basis function value at that point is 0.23 × 0.5 plus the weighted values ​​of the other seven basis functions.

[0032] One-dimensional convolution is performed on the standardized data using the generated dynamic wavelet basis function. Specifically, the dynamic wavelet basis function is used as the convolution kernel and convolved with the data sequence X to generate multi-level decomposition coefficients. Assuming the decomposition level is set to 5 levels, 5 sets of coefficients at different scales are obtained, each set containing approximation coefficients and detail coefficients.

[0033] The coefficients obtained from each decomposition layer are evaluated. For the approximation coefficients A_i of the i-th layer, their information entropy H_i is calculated. First, a histogram of the approximation coefficients is constructed, dividing the value range into 20 equally wide intervals. The probability of each interval is calculated, and then the information entropy is calculated using the entropy function. For example, the information entropy of the approximation coefficients of the 3rd layer is calculated as H_3 = 2.46. For the detail coefficients D_i, the ratio of their energy to the total energy E_i is calculated. The energy of the detail coefficients is defined as the sum of the squares of all coefficients, and the total energy is the sum of the energies of the detail coefficients of all layers. For example, the energy ratio of the detail coefficients of the 3rd layer is E_3 = 0.32.

[0034] Set the information entropy weight α = 0.6 and the energy proportion weight β = 0.4, and calculate the comprehensive evaluation index S_i = α × H_i + β × E_i for each layer. For example, the evaluation index for layer 3 is S_3 = 0.6 × 2.46 + 0.4 × 0.32 = 1.66. Compare the evaluation indices of all layers and select the layer with the largest index as the optimal decomposition layer. Assuming that the layer with the largest evaluation index among the 5 layers is layer 4, S_4 = 1.72, then the optimal decomposition layer is 4.

[0035] The standardized data is subjected to discrete wavelet transform using a defined dynamic wavelet basis function, decomposing it layer by layer according to the optimal decomposition level (4 levels) to obtain approximation coefficients and detail coefficients for each level. The discrete wavelet transform is implemented using the Mallat algorithm, including high-pass filtering and low-pass filtering operations. For example, for standardized data X, the first level decomposition yields approximation coefficients A_1 and detail coefficients D_1, the second level decomposition processes A_1 again to obtain A_2 and D_2, and so on.

[0036] Scale constraints are applied to the detail coefficients of each layer. For the detail coefficient D_j of layer j, the weighted average M_j of its value and the detail coefficients of its adjacent upper and lower layers is calculated. The weight of the upper layer is set to 0.4, and the weight of the lower layer is set to 0.6. For example, the weighted average M_2 of layer 2 is M_2 = 0.4 × D_1 + 0.6 × D_3. The scale constraint value C_j is obtained by subtracting the weighted average M_j from the detail coefficient D_j and then summing the squares. For example, C_2 = ∑(D_2 - M_2). 2 =0.087.

[0037] A preset scale threshold is set to 0.05. If the scale constraint value C_j is greater than this threshold, the detail coefficients need to be adjusted. The adjustment method is: D_j_new = D_j - λ(D_j - M_j), where λ is an adjustment factor, initially set to 0.2. This adjustment is repeated until the scale constraint value is less than the preset threshold. For example, after three adjustments, C_2 drops to 0.045, meeting the threshold requirement.

[0038] The adjusted detail coefficients are then convolved one-dimensionally using dynamic wavelet basis functions to reconstruct the signal, obtaining feature components at different frequency scales. The reconstruction process is the inverse of the discrete wavelet transform, implemented using a combination of upsampling and filters. For example, using the adjusted detail coefficients D_1_new from the first layer, high-frequency feature components are reconstructed, with a frequency range approximately 1 / 2 to 1 / 4 of the original signal sampling frequency; using the adjusted detail coefficients D_2_new from the second layer, mid-to-high-frequency feature components are reconstructed, with a frequency range approximately 1 / 4 to 1 / 8 of the original signal sampling frequency; and so on.

[0039] Through the above implementation method, the standardized data can be decomposed into multiple scales using wavelet transform to obtain feature components at different frequency scales, which provides a foundation for subsequent signal analysis and feature extraction.

[0040] In one optional implementation, a variational recurrent neural network is used to extract temporal features from the original feature sequence, and the temporal features are input into an autoencoder to generate a feature reconstruction sequence, including: Based on the feature vector of the current time step and the hidden state of the previous time step in the original feature sequence, the mean parameter and variance parameter of the variational posterior distribution are calculated. The variational posterior distribution is determined using the mean parameter and the variance parameter. The hidden variable of the current time step is sampled from the variational posterior distribution. The feature vector at the current time, the hidden state at the previous time, and the hidden variable at the current time are input into the variational recursive unit to obtain the hidden state at the current time. The variational posterior distribution calculation and the hidden state calculation are repeated until all feature vectors in the original feature sequence are processed to obtain the hidden state sequence corresponding to each time. The similarity between any two hidden states in the hidden state sequence is calculated to obtain a similarity score. The similarity score is normalized to obtain an attention weight sequence. The attention weight sequence is weighted and summed with the hidden state sequence at the corresponding time to obtain a context vector sequence. The hidden state sequence and the context vector sequence are concatenated along the temporal dimension and then input into the encoder to obtain initial temporal features. One-dimensional convolution operations with different kernel sizes are performed on the initial temporal features to obtain multiple scale features. The multiple scale features are weighted and summed with preset scale weights to obtain fused features. The fused features are input into the decoder to generate a feature reconstruction sequence.

[0041] In a method for extracting temporal features from an original feature sequence using a variational recurrent neural network, the original feature sequence to be processed is first obtained, which contains feature vectors at multiple time points. For each time point t in the original feature sequence, the mean and variance parameters of the variational posterior distribution are calculated based on the feature vector at the current time point t and the hidden state at the previous time point t-1. Specifically, the mean and variance parameters of the variational posterior distribution are obtained by nonlinear transformation of the feature vector at the current time point and the hidden state at the previous time point. For example, the feature vector and hidden state are mapped to an intermediate layer representation through a fully connected layer, and then the mean and variance parameters are obtained through two independent fully connected layers.

[0042] The variational posterior distribution is determined using the calculated mean and variance parameters. This distribution is typically set as a multidimensional Gaussian distribution, where the mean parameter is directly used as the distribution's mean, and the variance parameter is transformed using an exponential function to become the distribution's variance. Latent variables for the current time step are sampled from this variational posterior distribution. The sampling process employs a reparameterization technique: noise is sampled from a standard normal distribution, and then the noise is multiplied by the transformed variance parameter and added to the mean parameter to obtain the latent variables for the current time step.

[0043] The feature vector at the current time step, the hidden state at the previous time step, and the hidden variables at the current time step are input together into the variational recursive unit (NRU) to obtain the hidden state at the current time step. The NRU can be a variant based on a gated recursive unit (GRU) or a long short-term memory (LSTM) network. Taking LSTM as an example, the NRU controls the flow of information through input gates, forget gates, and output gates, and updates the unit state and hidden state by combining the hidden variables at the current time step. Through the above steps, the feature vectors at each time step in the original feature sequence are processed sequentially, ultimately obtaining the hidden state sequence corresponding to each time step.

[0044] After obtaining the hidden state sequence, the similarity between any two hidden states in the sequence is calculated to obtain a similarity score. Similarity calculation can use a dot product operation, that is, performing an inner product operation on the two hidden state vectors. For example, for a hidden state sequence of length T, T×T similarity scores will be calculated to form a similarity matrix. The similarity scores are then normalized to obtain the attention weight sequence. Normalization can use the softmax function to ensure that the sum of all weights is 1. The attention weight sequence is then weighted and summed with the hidden state sequence at the corresponding time step to obtain the context vector sequence. For each time step t, the attention weight at that time step is weighted and summed with the hidden states at all times to obtain the context vector at that time step.

[0045] The hidden state sequence and the context vector sequence are concatenated along the temporal dimension and then input into the encoder to obtain the initial temporal features. The concatenation operation merges the two sequences along the feature dimension, resulting in concatenated features that contain richer information. The encoder can be a multilayer perceptron, which maps the concatenated features to a new feature space to generate the initial temporal features.

[0046] By performing one-dimensional convolution operations with different kernel sizes on the initial temporal features, multiple scale features are obtained. For example, using one-dimensional convolutions with kernel sizes of 3, 5, and 7 can capture patterns across different time spans. Specifically, for a one-dimensional convolution with a kernel size of 3, 64 filters are used with a stride of 1, "same" padding, and the ReLU activation function; for a one-dimensional convolution with a kernel size of 5, 48 filters are used with the same parameters; and for a one-dimensional convolution with a kernel size of 7, 32 filters are used with the same parameters. Through these convolution operations at different scales, patterns of different granularities in temporal data can be captured.

[0047] Multiple scale features are weighted and summed with preset scale weights to obtain fused features. Scale weights can be preset fixed values ​​or learned parameters. For example, for three scale features, weights of 0.4, 0.3, and 0.3 can be set for weighted fusion. The fused features are then input into a decoder to generate a feature reconstruction sequence. The decoder can be a transposed convolutional network or a multilayer perceptron, which maps the fused features back to the dimensions of the original feature space, generating a feature reconstruction sequence of the same length as the original feature sequence.

[0048] In a specific example, the original feature sequence is 100 in length, and each feature vector has a dimension of 20. The latent state dimension of the variational recurrent unit is 64, and the latent variable dimension is 32. The encoder consists of two fully connected layers, mapping the input dimension from 128 (64 latent state dimension plus 64 context vector dimension) to 96 and 64, respectively. The convolutional operations use three different kernel sizes: 3, 5, and 7, producing scale feature dimensions of 64, 48, and 32, respectively. The scale weights are set to [0.4, 0.3, 0.3]. The decoder consists of two fully connected layers, mapping the fused feature dimension from 64 to 40 and 20, respectively. The last layer uses a tanh activation function to ensure the output value is within a reasonable range. In this way, the generated feature reconstruction sequence can effectively capture the temporal patterns in the original feature sequence.

[0049] In one optional implementation, the reconstruction error between the original feature sequence and the reconstructed sequence is calculated, and key time segments are selected based on the reconstruction error, including: The Euclidean distance between the original feature sequence and the reconstructed sequence at each time step is calculated to obtain the reconstruction error sequence. A time window of a preset size is set at each time step, and the reconstruction error values ​​within the time window are accumulated to obtain the local error value at that time step, thus forming a local error sequence. Each time step in the local error sequence is treated as a node in the graph structure. The feature vector corresponding to each node at that time step is extracted. The cosine similarity between the feature vectors of any two adjacent nodes is calculated to obtain the feature correlation coefficient. The feature correlation coefficient is then filled into the corresponding position of the initial edge weight matrix. The difference between the timestamps of any two adjacent nodes is calculated. The difference is then substituted into the Gaussian kernel function to obtain the time distance decay factor. The time distance decay factor is then filled into the corresponding position of the time weight matrix. The initial edge weight matrix is ​​multiplied by the corresponding element in the time weight matrix to obtain the final edge weight matrix. According to the edge weight matrix, each node obtains a local error value from its neighboring nodes, multiplies the local error value with the corresponding edge weight and sums them to obtain the update error value of the node, and calculates the update error values ​​of all nodes in turn to form an update error sequence. The local error sequence and the error values ​​at corresponding times in the updated error sequence are weighted and combined to obtain the cumulative error score sequence. The time segments in the cumulative error score sequence that are less than a preset score threshold are merged to obtain the final key time segments.

[0050] This embodiment provides a method for screening key time segments to identify abnormal or important time periods from the original feature sequence. The method first calculates the reconstruction error between the original feature sequence and the reconstructed sequence, and then filters out key time segments based on the error value and the time-space relationship.

[0051] In practical applications, the original feature sequence can be represented as a series of feature vectors corresponding to timestamps, such as multidimensional sensor data collected by an industrial device during continuous monitoring. A reconstructed sequence of the original feature sequence can be obtained through a pre-trained autoencoder or other reconstruction model. This reconstructed sequence reflects the main patterns and regularities of the original data, while the differences between the original and reconstructed sequences may indicate anomalies or critical events.

[0052] The reconstruction error is calculated as follows: For each time step t in the original feature sequence, the Euclidean distance between the original feature vector and the corresponding reconstructed feature vector is calculated. For example, if the original feature vector is [3.5, 2.1, 4.7] and the reconstructed feature vector is [3.2, 2.3, 4.5], then the reconstruction error at that time step is 0.42. After performing this operation on all time steps, the reconstruction error sequence [e1, e2, ..., en] is obtained.

[0053] To capture the clustering effect of errors over time, it is necessary to calculate local error values. Specifically, for each time t in the sequence, a time window of a preset size w is set, for example, w=5. The reconstruction error values ​​within the time window are summed to obtain the local error value at that time. For example, the local error value at time t is equal to the sum of the five reconstruction errors from time t-2 to t+2. In this way, a local error sequence [l1, l2, ..., ln] can be formed.

[0054] Next, a graph relating time segments is constructed, treating each moment in the local error sequence as a node in the graph. Graph construction involves two main steps: calculating the eigencorrelation between nodes and temporal distance decay.

[0055] For feature correlation calculation, the original feature vector ft corresponding to each node t at that time is extracted. The cosine similarity between the feature vectors fi and fj of any two adjacent nodes i and j is calculated to obtain the feature correlation coefficient sij. For example, if two feature vectors are [2.5, 3.1, 1.8] and [2.3, 2.9, 2.0], their cosine similarity is approximately 0.99. All calculated feature correlation coefficients are then filled into the corresponding positions of the initial edge weight matrix W0.

[0056] For time-distance decay calculation, for any two adjacent nodes i and j, calculate the difference Δt = |ti - tj| between their timestamps ti and tj. Substitute this difference into the Gaussian kernel function g(Δt) = exp(-(Δt)). 2 / σ 2 ), where σ is a preset parameter, such as σ=10. For example, if the timestamps of the two nodes are 125 and 130 respectively, then Δt=5. Substituting this into the Gaussian kernel function (assuming σ=10), the time distance decay factor is approximately 0.78. All calculated time distance decay factors are then filled into the corresponding positions of the time weight matrix Wt.

[0057] The final edge weight matrix W is obtained by element-wise multiplication, i.e., W[i][j] = W0[i][j] × Wt[i][j]. For example, if W0[i][j] = 0.99 and Wt[i][j] = 0.78, then W[i][j] = 0.77. This design makes the connections between nodes with similar features and close time intervals stronger.

[0058] Based on the constructed graph structure, error propagation and updates are performed. For each node i, the local error value lj is obtained from its neighboring node j. lj is multiplied by the corresponding edge weight W[i][j] and summed to obtain the update error value ui of that node. For example, if node i has three neighboring nodes j1, j2, and j3, with corresponding local error values ​​of 10.5, 8.2, and 12.7, and edge weights of 0.77, 0.65, and 0.82, then the update error value ui of node i is 10.5 × 0.77 + 8.2 × 0.65 + 12.7 × 0.82 = 26.62. The update error values ​​of all nodes are calculated sequentially to form the update error sequence [u1, u2, ..., un].

[0059] Finally, the local error sequence and the update error sequence are weighted and combined to obtain the cumulative error score sequence. Specifically, for each time t, st = α × lt + (1 - α) × ut is calculated, where α is the weighting coefficient, such as α = 0.6. This combination method considers both local error and global correlation information. For example, if the local error value at a certain time is 15.2 and the update error value is 26.62, then the cumulative error score is 15.2 × 0.6 + 26.62 × 0.4 = 19.73.

[0060] Each score in the cumulative error score sequence is compared with a preset threshold τ, such as τ=20.0. Scores below the threshold are marked as normal, and scores above the threshold are marked as abnormal. Consecutive abnormal moments are merged into a single time segment. For example, if scores from moments 125 to 130 all exceed the threshold, these six moments are merged into a single key time segment [125, 130]. ​​If the interval between two abnormal time segments is less than a preset merging threshold (e.g., 3 time units), these two segments are merged into a larger segment. The final set of all key time segments is the filtering result, for example, {[45, 52], [125, 138], [201, 209]}, indicating that abnormal or important events may have occurred within these time ranges.

[0061] The above method not only considers the reconstruction error of the data in the feature space, but also integrates the information of temporal continuity and feature correlation, which can more accurately screen out the real key time segments, reduce the interference of isolated error points, and improve the accuracy and robustness of anomaly detection.

[0062] In one optional implementation, the dynamic feature representation is input into a knowledge distillation network to extract start / stop decision rules, including: Calculate the query matrix, key matrix, and value matrix of the dynamic feature representation, multiply the query matrix by the transpose of the key matrix and divide by the square root of the feature dimension, input the calculation result into the normalized exponential function to obtain the attention weight, and multiply the attention weight by the value matrix to obtain the attention feature; Multiple temporal convolutional layers of different scales are constructed for the dynamic feature representation. Each temporal convolutional layer extracts feature patterns within different time windows to obtain convolutional features. The attention features and the convolutional features are fused with adaptive weights to obtain enhanced features. A dual-branch structure of teacher network and student network is constructed, and start / stop decision probability distributions corresponding to teacher network and student network are generated based on the enhanced features. Samples with the same start / stop label are constructed as positive sample pairs, and samples with different start / stop labels are constructed as negative sample pairs. The contrast loss between the positive sample pairs and the negative sample pairs is calculated. The KL divergence between the start / stop decision probability distribution of the teacher network and the start / stop decision probability distribution of the student network is used as the knowledge distillation loss. The knowledge distillation loss and the contrastive loss are weighted and combined to obtain the total loss function. Based on the total loss function, the gradient of each start-stop related feature in the student network with respect to the start-stop decision result is calculated to obtain the feature importance. The start / stop decision rules are scored based on a weighted combination of the importance of the features and the attention features, and start / stop decision rules with an importance score greater than a preset score threshold are extracted.

[0063] This invention relates to a method for extracting start-stop decision rules based on knowledge distillation networks. The method first inputs dynamic feature representations into a knowledge distillation network to extract start-stop decision rules.

[0064] In the attention mechanism computation, the system processes the dynamic feature representation, generating a query matrix, a key matrix, and a value matrix. For example, for a dynamic feature representation with a dimension of 256, three independent linear transformation layers generate a query matrix Q, a key matrix K, and a value matrix V of the same dimension. Multiplying the query matrix Q by the transpose of the key matrix K yields an attention score matrix, which represents the correlation strength between features at different time steps. To avoid the gradient vanishing problem caused by excessively large values, the product result is divided by the square root of the feature dimension (16 in this case). Subsequently, the calculation result is input into a normalized exponential function (Softmax) to obtain attention weights ranging from 0 to 1 with a sum of 1. Finally, these attention weights are multiplied by the value matrix V to obtain the attention features that capture long-term dependencies.

[0065] For temporal convolutional feature extraction, the system constructs multiple temporal convolutional layers at different scales to capture feature patterns within different time windows. Specifically, three temporal convolutional layers are used, employing one-dimensional convolutional kernels with window sizes of 3, 5, and 7, respectively, each containing 64 filters. These convolutional layers extract feature patterns within short-term (e.g., the last 3 time steps), medium-term (e.g., the last 5 time steps), and long-term (e.g., the last 7 time steps) time windows. For example, for historical device start-up and shutdown data, a short-term window might capture recent load change trends, a medium-term window might capture daily periodic patterns, and a long-term window might identify differences between weekdays and weekends. Through this multi-scale feature extraction, the system can comprehensively understand complex patterns in time-series data.

[0066] The feature fusion stage employs an adaptive weighting mechanism to fuse attention features and convolutional features. The system introduces a trainable weight parameter α (initially set to 0.5), and feature fusion is achieved by multiplying the attention feature by α × (1-α) × the convolutional feature. This adaptive fusion method allows the model to automatically adjust the importance ratio of the two features during training, ultimately resulting in an enhanced feature representation that includes both long-term dependencies and local temporal patterns.

[0067] In the design of the knowledge distillation network structure, a dual-branch structure of teacher and student networks was constructed. The teacher network adopts a more complex three-layer fully connected network (layer dimension 256-128-2), while the student network adopts a simple two-layer structure (layer dimension 128-2). Both networks generate probability distributions for start and stop decisions based on the aforementioned enhanced features. For example, for a feature input at a certain moment, the teacher network might output a probability distribution of [0.85, 0.15] (representing an 85% probability of starting and a 15% probability of stopping), while the student network might output a probability distribution of [0.82, 0.18].

[0068] To enhance the model's discriminative ability, the system constructs sample pairs based on contrastive learning. In the training data, samples with the same label (e.g., both being "start" decisions) are constructed as positive sample pairs, and samples with different labels are constructed as negative sample pairs. For each batch of data (batch size set to 64), the system randomly selects 16 positive sample pairs and 16 negative sample pairs. By calculating the difference between the cosine similarity of features in positive sample pairs and the cosine similarity of features in negative sample pairs, a contrastive loss function is constructed to encourage the model to bring the feature representations of similar decisions closer together and push the feature representations of different decisions further apart.

[0069] During knowledge distillation, the system calculates the KL divergence between the start / stop decision probability distributions of the teacher network and the student network as the knowledge distillation loss. This process essentially allows the student network to learn the "soft labels" of the teacher network, rather than just the hard labels (0 or 1). In this way, the rich knowledge of the teacher network is transferred to the lighter-weight student network. In actual calculations, if the teacher network outputs [0.85, 0.15] and the student network outputs [0.82, 0.18], the KL divergence value is approximately 0.0026, indicating that the two distributions are relatively close.

[0070] The total loss function is constructed using a weighted combination approach, integrating the knowledge distillation loss and the contrastive loss. Specifically, the total loss = β × knowledge distillation loss + (1-β) × contrastive loss, where β is a weight hyperparameter set to 0.7 to balance the contributions of the two losses. Based on this total loss function, the system calculates the gradient of each start / stop-related feature in the student network with respect to the start / stop decision outcome using the backpropagation algorithm, thus obtaining the feature importance score.

[0071] Finally, the system assigns an importance score to the start / stop decision rules based on a weighted combination of feature importance and attention features. The specific calculation formula is: Rule Importance = γ × Feature Importance + (1-γ) × Attention Weight, where γ is set to 0.6. A scoring threshold of 0.75 (within the range of 0 to 1) is set, and the system extracts start / stop decision rules with an importance score greater than this threshold as the final result. For example, in industrial equipment start / stop scenarios, the system might identify a high-value decision rule such as "the equipment should stop operating when the load is below 30% for three consecutive hours and the current time is during peak electricity consumption periods."

[0072] Through the above method, the present invention realizes the function of automatically extracting high-value start-stop decision rules from historical optimization experience, providing an interpretable optimization strategy for the efficient operation of industrial equipment.

[0073] In one optional implementation, a similarity score between the current decision and historical experience is calculated using a contrastive learning method, and the learning direction of the policy network is adjusted based on the similarity score, including: Using a sliding time window, temporal feature sequences are extracted sequentially from the historical experience; the temporal feature sequences are input into the forward and backward networks of a bidirectional long short-term memory network to generate forward hidden state sequences and backward hidden state sequences, respectively; the forward hidden state sequences and the backward hidden state sequences are concatenated along the feature dimension to obtain a temporal representation vector. For the time-series representation vector, time-series pooling operations are performed at different time scales to obtain multi-scale feature vectors. The similarity between different feature vectors is calculated. Feature vector pairs with similarity greater than a preset positive sample threshold are stored as positive sample pairs in the positive sample set, and feature vector pairs with similarity less than a preset negative sample threshold are stored as negative sample pairs in the negative sample set; wherein, the preset positive sample threshold is greater than the preset negative sample threshold. The multi-scale feature vectors are projected into the state space and the action space respectively. The cosine distance between the feature vectors is calculated in the state space to obtain the state similarity, and the cosine distance between the feature vectors is calculated in the action space to obtain the action similarity. The state similarity and the action similarity are concatenated and input into the attention network. The attention network outputs dynamic weight coefficients. The state similarity and the action similarity are weighted and summed according to the dynamic weight coefficients to obtain a similarity score. Based on the similarity score, the sample pair with the highest similarity is selected from the set of positive samples as reference experience, and the action strategy of the reference experience is used as the learning direction for the current decision.

[0074] This embodiment provides a method for calculating the similarity score between the current decision and historical experience using a contrastive learning approach, and adjusting the learning direction of the policy network based on this similarity score. This method improves the accuracy and efficiency of decision-making by calculating the similarity between the current decision and historical experience, selecting the most similar reference experience from historical experience, and using its action policy as the learning direction for the current decision.

[0075] In practical applications, the system first uses a sliding time window to sequentially extract time-series feature sequences from historical experience. For example, for an autonomous driving system, the sliding window size can be set to 10 seconds with a step size of 1 second to extract time-series feature sequences containing information such as vehicle speed, steering wheel angle, and the positions of surrounding vehicles from historical driving data. Specifically, at time point t, the system extracts the feature sequence [ft-9, ft-8, ..., ft] from the time period t-9 to t, where ft represents the feature vector at time point t, containing information such as vehicle speed of 60 km / h, steering wheel angle of 15 degrees, and distance to the vehicle ahead of 20 meters.

[0076] Next, the system inputs the extracted temporal feature sequence into a Bidirectional Long Short-Term Memory (BiLSTM) network to generate forward and backward hidden state sequences. The BiLSTM network consists of two parts: a forward LSTM and a backward LSTM. The forward LSTM processes the sequence from the beginning, generating the forward hidden state sequence [hf_t-9, hf_t-8, ..., hf_t]; the backward LSTM processes the sequence from the end, generating the backward hidden state sequence [hb_t, hb_t-1, ..., hb_t-9]. Each hidden state vector has a dimension of 128. The system concatenates the forward and backward hidden state sequences along the feature dimensions to obtain the temporal representation vector [h_t-9, h_t-8, ..., h_t], where h_t = [hf_t; hb_t], with a dimension of 256.

[0077] For the obtained temporal representation vectors, the system performs temporal pooling operations at different time scales to obtain multi-scale feature vectors. Specifically, the system sets three different pooling window sizes: 2, 5, and 10, corresponding to short-term, medium-term, and long-term time scales, respectively. For pooling operations with a window size of 2, the system averages the representation vectors of two adjacent time steps to obtain feature vector z1; for pooling operations with a window size of 5, the system averages the representation vectors of five consecutive time steps to obtain feature vector z2; for pooling operations with a window size of 10, the system averages the representation vectors of the entire sequence to obtain feature vector z3. In this way, the system obtains a multi-scale feature vector set Z = {z1, z2, z3}.

[0078] The system calculates the similarity between different feature vectors. Feature vector pairs with a similarity greater than a preset positive sample threshold are stored as positive sample pairs in the positive sample set, while feature vector pairs with a similarity less than a preset negative sample threshold are stored as negative sample pairs in the negative sample set. In practical applications, the system sets the preset positive sample threshold to 0.8 and the preset negative sample threshold to 0.3. For example, if the similarity between feature vectors z1 and z2 is calculated to be 0.85, which is greater than the preset positive sample threshold of 0.8, then (z1, z2) is stored as a positive sample pair in the positive sample set. Similarly, if the similarity between feature vectors z1 and z3 is calculated to be 0.25, which is less than the preset negative sample threshold of 0.3, then (z1, z3) is stored as a negative sample pair in the negative sample set.

[0079] The system projects multi-scale feature vectors into the state space and action space, respectively. Specifically, the system uses two fully connected layers as projection functions. The feature vector z is projected into the state space to obtain the state representation s = Ws*z + bs, with a dimension of 128; the feature vector z is projected into the action space to obtain the action representation a = Wa*z + ba, with a dimension of 64. Here, Ws and Wa are the projection matrices of the state space and action space, respectively, and bs and ba are the corresponding bias vectors.

[0080] In the state space, the system calculates the cosine distance between feature vectors to obtain state similarity. For two state representations s1 and s2, the state similarity is calculated as the dot product of s1 and s2 divided by the product of their Euclidean norms. For example, for state representations s1=[0.5, 0.3, 0.8, ...] and s2=[0.4, 0.6, 0.7, ...], the calculated state similarity is 0.92. Similarly, in the action space, the system calculates the cosine distance between feature vectors to obtain action similarity, for example, an action similarity of 0.78.

[0081] The system concatenates state similarity and action similarity and inputs the result into an attention network, which then outputs dynamic weight coefficients. The attention network consists of a fully connected layer and a softmax layer, with inputs of [0.92, 0.78] and outputs dynamic weight coefficients of [0.6, 0.4]. Based on these dynamic weight coefficients, the system performs a weighted sum of the state and action similarities, resulting in a similarity score of 0.92*0.6 + 0.78*0.4 = 0.864.

[0082] Based on similarity scores, the system selects the pair of samples with the highest similarity from the positive sample set as reference experience. For example, if the positive sample set contains sample pairs (z1, z2), (z2, z3), and (z1, z4) with similarity scores of 0.864, 0.752, and 0.793 respectively, the system selects the sample pair (z1, z2) with the highest similarity score as reference experience. The system uses the action policy from the reference experience as the learning direction for the current decision. Specifically, it uses the action representation a2 from the reference experience as the learning target of the policy network, updating the parameters of the policy network using gradient descent to make the action representation a1 of the current decision move closer to a2. In this way, the system can learn effective decision-making strategies from historical experience, improving the accuracy and efficiency of decision-making.

[0083] A second aspect of the present invention provides an intelligent start-stop control system for a concentrated solar power (CSP) system based on a DCS, comprising: The first unit is used to collect real-time operating data of the solar thermal power generation system through the distributed control system (DCS); to standardize the real-time operating data; to decompose the standardized data into multiple scales using wavelet transform to obtain feature components at different frequency scales; to reduce the dimensionality of the feature components using principal component analysis to obtain the principal feature vector; and to combine the principal feature vector with the real-time operating data to form the original feature sequence. The second unit is used to extract the temporal features of the original feature sequence using a variational recurrent neural network, input the temporal features into an autoencoder to generate a feature reconstruction sequence; calculate the reconstruction error between the original feature sequence and the reconstruction sequence, select key time segments based on the reconstruction error; and input the key time segments into a recurrent neural network to generate a dynamic feature representation of the system state. The third unit is used to input the dynamic feature representation into a knowledge distillation network to extract start-stop decision rules; construct a policy network for generating start-stop control strategies, wherein the policy network adopts a deep neural network structure, and encodes the start-stop decision rules as the initial parameters of the policy network; calculate the similarity score between the current decision and historical experience using a contrastive learning method, and adjust the learning direction of the policy network based on the similarity score; The fourth unit is used to deploy the adjusted strategy network to the field control unit of the DCS. The field control unit sends start-stop control commands to the solar thermal power generation system according to the strategy output by the strategy network to realize intelligent start-stop control of the system.

[0084] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0085] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0086] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent start-stop control of a concentrated solar power (CSP) system based on DCS, characterized in that, include: Real-time operating data of the solar thermal power generation system is collected through a distributed control system (DCS). The real-time operational data is then standardized. Wavelet transform is used to perform multi-scale decomposition on the standardized data to obtain feature components at different frequency scales. Principal component analysis is used to reduce the dimensionality of the feature components to obtain the principal feature vectors. The principal feature vectors are then combined with the real-time running data to form the original feature sequence. The process involves using wavelet transform to decompose standardized data at multiple scales to obtain feature components at different frequency scales. This includes: calculating local statistical features based on a sliding time window; feeding these local statistical features into a trained neural network to obtain weight coefficients for each basis function; multiplying these weight coefficients by each basis function in a preset set and summing the results to obtain dynamic wavelet basis functions; performing one-dimensional convolution on the standardized data using the dynamic wavelet basis functions to generate multi-level decomposition coefficients; calculating the information entropy of each approximate coefficient using an entropy function; calculating the ratio of the energy of the detail coefficients to the total energy to obtain the energy percentage; and comparing the information entropy with the energy percentage. The evaluation index of each layer is obtained by multiplying each layer by its corresponding preset weight and then summing the results. The layer with the highest evaluation index is determined as the optimal decomposition layer. The standardized data is then subjected to discrete wavelet transform using the dynamic wavelet basis function. The data is decomposed layer by layer according to the optimal decomposition layer to obtain the approximation coefficients and detail coefficients of each layer. The weighted average of the detail coefficients of each layer and the detail coefficients of its adjacent upper and lower layers is calculated. The sum of squares after subtracting the weighted average from the detail coefficients of that layer is obtained to obtain the scale constraint value. The value of the detail coefficients is adjusted according to the scale constraint value until the scale constraint value is less than a preset scale threshold. The adjusted detail coefficients are then subjected to one-dimensional convolution operation using the dynamic wavelet basis function to reconstruct the feature components at different frequency scales. A variational recurrent neural network is used to extract temporal features from the original feature sequence. These temporal features are then input into an autoencoder to generate a feature reconstruction sequence. The reconstruction error between the original feature sequence and the reconstructed sequence is calculated, and key time segments are selected based on the reconstruction error. The key time segments are then input into a recurrent neural network to generate a dynamic feature representation of the system state. Calculating the reconstruction error between the original feature sequence and the reconstructed sequence, and selecting key time segments based on the reconstruction error, includes: calculating the Euclidean distance between the original feature sequence and the reconstructed sequence at each time step to obtain a reconstruction error sequence; setting a time window of a preset size at each time step; accumulating the reconstruction error values ​​within the time window to obtain the local error value at that time step, forming a local error sequence; treating each time step in the local error sequence as a node in a graph structure; extracting the feature vector corresponding to each node at that time step; and calculating the difference between the feature vectors of any two adjacent nodes. The cosine similarity is used to obtain the feature correlation coefficient, which is then filled into the corresponding position of the initial edge weight matrix. The difference between the timestamps of any two adjacent nodes is calculated, and the difference is substituted into the Gaussian kernel function to obtain the time distance decay factor. The time distance decay factor is then filled into the corresponding position of the time weight matrix. The initial edge weight matrix is ​​multiplied by the corresponding element in the time weight matrix to obtain the final edge weight matrix. According to the edge weight matrix, each node obtains a local error value from its adjacent nodes, multiplies the local error value by the corresponding edge weight, and sums the results to obtain the update error value of that node. The update error values ​​of all nodes are calculated sequentially to form an update error sequence. The local error sequence and the error values ​​at corresponding times in the update error sequence are weighted and combined to obtain a cumulative error score sequence. Time segments in the cumulative error score sequence that are less than a preset score threshold are merged to obtain the final key time segments. The dynamic feature representation is input into a knowledge distillation network to extract start-stop decision rules; a policy network for generating start-stop control strategies is constructed, the policy network adopts a deep neural network structure, and the start-stop decision rules are encoded as the initial parameters of the policy network; a contrastive learning method is used to calculate the similarity score between the current decision and historical experience, and the learning direction of the policy network is adjusted based on the similarity score; The adjusted strategy network is deployed to the field control unit of the DCS. The field control unit sends start-stop control commands to the solar thermal power generation system according to the strategy output by the strategy network, thereby realizing intelligent start-stop control of the system.

2. The method according to claim 1, characterized in that, The temporal features of the original feature sequence are extracted using a variational recurrent neural network, and the temporal features are input into an autoencoder to generate a feature reconstruction sequence, including: Based on the feature vector of the current time step and the hidden state of the previous time step in the original feature sequence, the mean parameter and variance parameter of the variational posterior distribution are calculated. The variational posterior distribution is determined using the mean parameter and the variance parameter. The hidden variable of the current time step is sampled from the variational posterior distribution. The feature vector at the current time, the hidden state at the previous time, and the hidden variable at the current time are input into the variational recursive unit to obtain the hidden state at the current time. The variational posterior distribution calculation and the hidden state calculation are repeated until all feature vectors in the original feature sequence are processed to obtain the hidden state sequence corresponding to each time. The similarity between any two hidden states in the hidden state sequence is calculated to obtain a similarity score. The similarity score is normalized to obtain an attention weight sequence. The attention weight sequence is weighted and summed with the hidden state sequence at the corresponding time to obtain a context vector sequence. The hidden state sequence and the context vector sequence are concatenated along the temporal dimension and then input into the encoder to obtain initial temporal features. One-dimensional convolution operations with different kernel sizes are performed on the initial temporal features to obtain multiple scale features. The multiple scale features are weighted and summed with preset scale weights to obtain fused features. The fused features are input into the decoder to generate a feature reconstruction sequence.

3. The method according to claim 1, characterized in that, The dynamic feature representation is input into a knowledge distillation network to extract start-stop decision rules, including: Calculate the query matrix, key matrix, and value matrix of the dynamic feature representation, multiply the query matrix by the transpose of the key matrix and divide by the square root of the feature dimension, input the calculation result into the normalized exponential function to obtain the attention weight, and multiply the attention weight by the value matrix to obtain the attention feature; Multiple temporal convolutional layers of different scales are constructed for the dynamic feature representation. Each temporal convolutional layer extracts feature patterns within different time windows to obtain convolutional features. The attention features and the convolutional features are fused with adaptive weights to obtain enhanced features. A dual-branch structure of teacher network and student network is constructed, and start / stop decision probability distributions corresponding to teacher network and student network are generated based on the enhanced features. Samples with the same start / stop label are constructed as positive sample pairs, and samples with different start / stop labels are constructed as negative sample pairs. The contrast loss between the positive sample pairs and the negative sample pairs is calculated. The KL divergence between the start / stop decision probability distribution of the teacher network and the start / stop decision probability distribution of the student network is used as the knowledge distillation loss. The knowledge distillation loss and the contrastive loss are weighted and combined to obtain the total loss function. Based on the total loss function, the gradient of each start-stop related feature in the student network with respect to the start-stop decision result is calculated to obtain the feature importance. The start / stop decision rules are scored based on a weighted combination of the importance of the features and the attention features, and start / stop decision rules with an importance score greater than a preset score threshold are extracted.

4. The method according to claim 1, characterized in that, The similarity score between the current decision and historical experience is calculated using a contrastive learning method. The learning direction of the policy network is then adjusted based on this similarity score, including: Using a sliding time window, temporal feature sequences are extracted sequentially from the historical experience; the temporal feature sequences are input into the forward and backward networks of a bidirectional long short-term memory network to generate forward hidden state sequences and backward hidden state sequences, respectively; the forward hidden state sequences and the backward hidden state sequences are concatenated along the feature dimension to obtain a temporal representation vector. For the time-series representation vector, time-series pooling operations are performed at different time scales to obtain multi-scale feature vectors. The similarity between different feature vectors is calculated. Feature vector pairs with similarity greater than a preset positive sample threshold are stored as positive sample pairs in the positive sample set, and feature vector pairs with similarity less than a preset negative sample threshold are stored as negative sample pairs in the negative sample set; wherein, the preset positive sample threshold is greater than the preset negative sample threshold. The multi-scale feature vectors are projected into the state space and the action space respectively. The cosine distance between the feature vectors is calculated in the state space to obtain the state similarity, and the cosine distance between the feature vectors is calculated in the action space to obtain the action similarity. The state similarity and the action similarity are concatenated and input into the attention network. The attention network outputs dynamic weight coefficients. The state similarity and the action similarity are weighted and summed according to the dynamic weight coefficients to obtain a similarity score. Based on the similarity score, the sample pair with the highest similarity is selected from the set of positive samples as reference experience, and the action strategy of the reference experience is used as the learning direction for the current decision.

5. A DCS-based intelligent start-stop control system for a concentrated solar power (CSP) system, used to implement the method as described in any one of claims 1-4, characterized in that, include: The first unit is used to collect real-time operating data of the solar thermal power generation system through the distributed control system (DCS). The real-time operational data is then standardized. Wavelet transform is used to perform multi-scale decomposition on the standardized data to obtain feature components at different frequency scales. Principal component analysis is used to reduce the dimensionality of the feature components to obtain the principal feature vectors. The principal feature vectors are then combined with the real-time running data to form the original feature sequence. The second unit is used to extract the temporal features of the original feature sequence using a variational recurrent neural network, input the temporal features into an autoencoder to generate a feature reconstruction sequence; calculate the reconstruction error between the original feature sequence and the reconstruction sequence, select key time segments based on the reconstruction error; and input the key time segments into a recurrent neural network to generate a dynamic feature representation of the system state. The third unit is used to input the dynamic feature representation into a knowledge distillation network to extract start-stop decision rules; construct a policy network for generating start-stop control strategies, wherein the policy network adopts a deep neural network structure, and encodes the start-stop decision rules as the initial parameters of the policy network; calculate the similarity score between the current decision and historical experience using a contrastive learning method, and adjust the learning direction of the policy network based on the similarity score; The fourth unit is used to deploy the adjusted strategy network to the field control unit of the DCS. The field control unit sends start-stop control commands to the solar thermal power generation system according to the strategy output by the strategy network to realize intelligent start-stop control of the system.

6. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 4.

Images