Phase change energy storage heat reservoir anomaly detection system and method based on deep learning

Abstract

The invention relates to the field of intelligent detection, and provides a phase change energy storage heat reservoir anomaly detection system and method based on deep learning, and the method comprises the steps: carrying out the independent time sequence coding of the PCM temperature, the inlet and outlet temperature, and the environment temperature, so as to extract the time correlation characteristics of all parameters; performing bitwise difference operation on the environment temperature time sequence correlation implicit features and inlet and outlet temperature time sequence correlation implicit features to strip noise interference of environment temperature fluctuation on a heat transfer link from a feature space, and constructing a thermodynamic response model between the inlet-outlet temperature and the PCM temperature by using chain reasoning coding; pCM temperature simulation features based on normal working conditions are generated, and finally, thermodynamic constitutive relation deviation degree detection after environmental noise decoupling is achieved by comparing difference vectors of the simulation features and actual features. Therefore, the detection sensitivity of the abnormity of the phase change energy storage heat reservoir can be obviously improved.

Description

Technical Field

[0001] This application relates to the field of intelligent detection, and more specifically, to an abnormal detection system and method for a phase change energy storage thermal reservoir based on deep learning. Background Art

[0002] As an important part of efficient thermal energy management, the phase change energy storage thermal reservoir realizes the storage and release of energy through the latent heat characteristics of phase change materials (PCM), and has wide applications in fields such as industrial waste heat recovery and building energy conservation. During operation, there are complex dynamic coupling relationships among parameters such as the temperature of PCM inside the thermal reservoir, the inlet and outlet temperatures of the fluid, and the ambient temperature. Factors such as equipment aging, performance attenuation of phase change materials, or sudden changes in the external environment may cause thermodynamic imbalance, resulting in a decrease in energy storage efficiency or even system failure. Therefore, real-time abnormal detection of the thermal reservoir operation state is a key technical link to ensure the safety and economy of the system.

[0003] Traditional abnormal detection methods mainly rely on threshold-based statistical monitoring or traditional statistical models, but there are significant defects in practical applications: First, the time-series data collected by multi-source sensors of the phase change energy storage system (such as inlet temperature, outlet temperature, PCM temperature, etc.) are affected by equipment vibration, electromagnetic interference, and environmental fluctuations, showing high-noise characteristics, and traditional filtering algorithms are difficult to effectively separate effective signals from random noise; Second, the non-linear dynamic correlations among temperature parameters are not fully modeled, and existing methods often analyze single parameters in isolation or adopt linear correlation assumptions, resulting in incomplete extraction of abnormal features; Third, the dynamic influence of the ambient temperature as an external interference source is not explicitly decoupled, making it easy for real abnormal signals to be masked by environmental noise. These problems together lead to high false alarm rates and low sensitivity of traditional detection models, and it is difficult to meet the accurate monitoring requirements under complex working conditions.

[0004] Therefore, an abnormal detection scheme for a phase change energy storage thermal reservoir based on deep learning is desired. Summary of the Invention

[0005] This application aims at the deficiencies in the prior art and provides an abnormal detection system and method for a phase change energy storage thermal reservoir based on deep learning.

[0006] According to one aspect of this application, an abnormal detection method for a phase change energy storage thermal reservoir based on deep learning is provided, which includes:

[0007] Extracting a time-series data set of operation data from the monitoring system of the phase change energy storage thermal reservoir, where the operation data includes PCM temperature, inlet temperature, outlet temperature, and ambient temperature;

[0008] Group the time series dataset of the operation data and perform time series encoding to obtain an inlet temperature time series correlation implicit feature encoding vector, an outlet temperature time series correlation implicit feature encoding vector, an ambient temperature time series correlation implicit feature encoding vector, and a PCM temperature time series correlation implicit feature encoding vector;

[0009] Based on the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector, obtain the PCM temperature simulation inference time series correlation implicit feature, including: performing PCM temperature simulation inference based on noise-reduced inlet and outlet temperature time series guidance on the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector to obtain the PCM temperature simulation inference time series correlation implicit feature;

[0010] Based on the PCM temperature simulation inference time series correlation implicit feature and the PCM temperature time series correlation implicit feature encoding vector, determine whether there is an abnormality in the phase change energy storage heat reservoir.

[0011] According to another aspect of the present application, there is provided an abnormal detection system for a phase change energy storage heat reservoir based on deep learning, which includes:

[0012] An operation data extraction module, configured to extract a time series dataset of operation data from the monitoring system of the phase change energy storage heat reservoir, where the operation data includes PCM temperature, inlet temperature, outlet temperature, and ambient temperature;

[0013] A data encoding and grouping module, configured to group the time series dataset of the operation data and perform time series encoding to obtain an inlet temperature time series correlation implicit feature encoding vector, an outlet temperature time series correlation implicit feature encoding vector, an ambient temperature time series correlation implicit feature encoding vector, and a PCM temperature time series correlation implicit feature encoding vector;

[0014] A PCM temperature simulation inference module, configured to obtain the PCM temperature simulation inference time series correlation implicit feature based on the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector, where the PCM temperature simulation inference module is configured to: perform PCM temperature simulation inference based on noise-reduced inlet and outlet temperature time series guidance on the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector to obtain the PCM temperature simulation inference time series correlation implicit feature;

[0015] Anomaly determination module, configured to determine whether there is an anomaly in the phase change energy storage heat reservoir based on the PCM temperature simulation inference time series associated implicit feature and the PCM temperature time series associated implicit feature encoding vector.

[0016] Due to the adoption of the above technical solution, the present application has significant technical effects:

[0017] The phase change energy storage heat reservoir anomaly detection system and method based on deep learning provided by the present application first performs independent time series encoding on the PCM temperature, inlet and outlet temperatures, and ambient temperature to extract the time correlation features of each parameter, and then performs bitwise difference operation on the ambient temperature time series associated implicit feature and the inlet and outlet temperature time series associated implicit features to strip the noise interference of the ambient temperature fluctuation on the heat transfer link from the feature space. Then, a thermodynamic response model between the inlet-outlet temperature and the PCM temperature is constructed by using chain inference encoding to generate the PCM temperature simulation features based on the normal working conditions. Finally, by comparing the difference vector between the simulation features and the actual features, the deviation degree detection of the thermodynamic constitutive relationship after decoupling the environmental noise is realized. In this way, the detection sensitivity of the phase change energy storage heat reservoir anomaly can be significantly improved. Description of the Drawings

[0018] By describing the embodiments of the present application in more detail in conjunction with the drawings, the above and other objects, features, and advantages of the present application will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application, and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 It is a flowchart of the phase change energy storage heat reservoir anomaly detection method based on deep learning according to an embodiment of the present application.

[0020] Figure 2 It is a flowchart of step S2 in the phase change energy storage heat reservoir anomaly detection method based on deep learning according to an embodiment of the present application.

[0021] Figure 3 It is a flowchart of step S3 in the phase change energy storage heat reservoir anomaly detection method based on deep learning according to an embodiment of the present application.

[0022] Figure 4 It is a flowchart of step S32 in the phase change energy storage heat reservoir anomaly detection method based on deep learning according to an embodiment of the present application.

[0023] Figure 5 It is a flowchart of step S32-2 in the phase change energy storage heat reservoir anomaly detection method based on deep learning according to an embodiment of the present application.

[0024] Figure 6 It is a flowchart of step S4 in the abnormal detection method of the phase change energy storage thermal reservoir based on deep learning according to an embodiment of the present application.

[0025] Figure 7 It is a system block diagram of the abnormal detection system of the phase change energy storage thermal reservoir based on deep learning according to an embodiment of the present application. Detailed implementation manners

[0026] Next, exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0027] The phase change energy storage thermal reservoir is an important part of efficient thermal energy management. It utilizes the latent heat characteristics of phase change materials (PCMs) to achieve energy storage and release, and is widely used in fields such as industrial waste heat recovery and building energy conservation. When the thermal reservoir operates, there are complex dynamic coupling relationships among parameters such as the temperature of the internal PCM, the inlet and outlet temperatures of the fluid, and the ambient temperature. Equipment aging, material property attenuation, or external environment mutations will cause thermodynamic imbalance, reduce the energy storage efficiency and even lead to system failures. Therefore, real-time abnormal detection is crucial for ensuring system safety and economy.

[0028] Traditional abnormal detection methods have significant defects: First, the time-series data collected by multi-source sensors in the phase change energy storage system has high noise, and traditional filtering algorithms are difficult to effectively separate signals from noise; second, the nonlinear dynamic correlations among temperature parameters are not fully modeled, and single parameters are analyzed in isolation or linear correlation assumptions are used, resulting in incomplete extraction of abnormal features; third, the dynamic influence of the ambient temperature is not explicitly decoupled, and real abnormal signals are easily masked by ambient noise. These problems lead to high false alarm rates and low sensitivities of traditional detection models, making it difficult to meet the accurate monitoring requirements under complex working conditions.

[0029] To address the above technical problems, the technical concept of this application is to use deep learning-based data analysis and processing techniques to first perform independent time series encoding on the PCM temperature, inlet and outlet temperatures, and ambient temperature, and extract the time correlation features of each parameter; then perform a bitwise difference operation on the ambient temperature time series correlation hidden features and the inlet and outlet temperature time series correlation hidden features to strip the noise interference of ambient temperature fluctuations on the heat transfer link from the feature space; then use chain reasoning encoding to construct a thermodynamic response model between the inlet-outlet temperature and the PCM temperature, and generate PCM temperature simulation features based on normal operating conditions; finally, by comparing the difference vector between the simulation features and the actual features, the deviation degree detection of the thermodynamic constitutive relationship after ambient noise decoupling is realized. This solution breakthroughly uses the ambient temperature as an interpretable interference source for feature space compensation, and at the same time implicitly models the multi-parameter non-linear coupling relationship through deep learning, which can effectively suppress the interference of sensor noise and operating condition fluctuations on anomaly discrimination, and significantly improve the detection sensitivity of latent faults such as phase change material performance decay and heat transfer path blockage.

[0030] Figure 1 FIG. is a flowchart of a deep learning-based abnormal detection method for a phase change energy storage thermal reservoir according to an embodiment of the present application. As Figure 1 shown, the deep learning-based abnormal detection method for a phase change energy storage thermal reservoir according to an embodiment of the present application includes: S1, extracting a time series data set of operation data from the monitoring system of the phase change energy storage thermal reservoir, where the operation data includes PCM temperature, inlet temperature, outlet temperature, and ambient temperature; S2, performing data grouping and time series encoding on the time series data set of the operation data to obtain an inlet temperature time series correlation hidden feature encoding vector, an outlet temperature time series correlation hidden feature encoding vector, an ambient temperature time series correlation hidden feature encoding vector, and a PCM temperature time series correlation hidden feature encoding vector; S3, based on the inlet temperature time series correlation hidden feature encoding vector, the outlet temperature time series correlation hidden feature encoding vector, and the ambient temperature time series correlation hidden feature encoding vector, obtaining a PCM temperature simulation inference time series correlation hidden feature; S4, based on the PCM temperature simulation inference time series correlation hidden feature and the PCM temperature time series correlation hidden feature encoding vector, determining whether there is an abnormality in the phase change energy storage thermal reservoir.

[0031] In step S1, a time series dataset of operation data is extracted from the monitoring system of the phase change energy storage thermal reservoir. The operation data includes PCM temperature, inlet temperature, outlet temperature, and ambient temperature. It should be understood that the PCM temperature time series reflects the energy storage and release states of the phase change material. Its abnormal fluctuations or deviations from the normal phase change temperature range may directly indicate the performance degradation of the phase change material or the decline in the internal heat transfer efficiency of the thermal reservoir. The inlet temperature and outlet temperature time series reflect the heat exchange process between the fluid and the phase change material. The temperature difference change between the two and the dynamic response over time can reflect the energy input and output efficiency of the thermal reservoir. If the outlet temperature abnormally decreases at the same inlet temperature, it may mean that the heat exchange channel is blocked or the fluid flow rate is abnormal. The ambient temperature time series, as an external disturbance variable, its change will affect the heat loss rate between the thermal reservoir and the outside world. If the PCM temperature does not show the expected coupled change when the ambient temperature fluctuates (such as the PCM temperature abnormally rapidly decreasing when the ambient temperature suddenly drops), it may imply the failure of the thermal reservoir insulation layer or abnormal external heat dissipation. In summary, the time series dataset of the operation data is not only a digital mapping of the system operation state, but also the core basis for detecting anomalies by mining the dynamic coupling relationship between parameters. The information such as the time correlation, the non-linear association between parameters, and the influence of external disturbances contained in it can provide multi-dimensional feature inputs for the anomaly detection of the phase change energy storage thermal reservoir, enabling the system to identify complex anomaly patterns that are difficult to capture by traditional threshold methods through a data-driven approach.

[0032] In step S2, the time series dataset of the operation data is grouped and time series encoded to obtain an inlet temperature time series associated hidden feature encoding vector, an outlet temperature time series associated hidden feature encoding vector, an ambient temperature time series associated hidden feature encoding vector, and a PCM temperature time series associated hidden feature encoding vector. Specifically, Figure 2 The flowchart of step S2 in the method for anomaly detection of a phase change energy storage thermal reservoir based on deep learning according to an embodiment of the present application. As Figure 2 shown, the step S2 includes: S21, grouping the time series dataset of the operation data to obtain a PCM temperature time series, an inlet temperature time series, an outlet temperature time series, and an ambient temperature time series; S22, performing time series encoding on the PCM temperature time series, the inlet temperature time series, the outlet temperature time series, and the ambient temperature time series to obtain the PCM temperature time series associated hidden feature encoding vector, the inlet temperature time series associated hidden feature encoding vector, the outlet temperature time series associated hidden feature encoding vector, and the ambient temperature time series associated hidden feature encoding vector.

[0033] In step S21, the time series dataset of the operating data is grouped to obtain a PCM temperature time series, an inlet temperature time series, an outlet temperature time series, and an ambient temperature time series. Correspondingly, considering that although the inlet temperature, outlet temperature, PCM temperature, and ambient temperature parameters collected by multi-source sensors have physical correlations, there are essential differences in the thermodynamic process dimensions they represent. The inlet temperature reflects the energy input state of the heat transfer fluid, the outlet temperature characterizes the energy release efficiency of the system, the PCM temperature is directly related to the energy storage characteristics of the phase change material, and the ambient temperature, as an external disturbance variable, is independent of the heat transfer link inside the heat reservoir. When traditional methods mix and process multi-modal temperature parameters, the time evolution laws of different physical processes will be blurred. For example, the coupling effect between ambient temperature fluctuations and inlet temperature changes may mask real heat transfer anomalies. In addition, the differences in the sampling frequencies, noise sources, and time lag characteristics of different parameters are not distinguished, and direct modeling easily leads to feature confusion. Based on this, the present application groups the time series dataset of the operating data to obtain a PCM temperature time series, an inlet temperature time series, an outlet temperature time series, and an ambient temperature time series. In this way, by separately extracting the time series of the PCM temperature, inlet temperature, outlet temperature, and ambient temperature, it can be ensured that the time-dependent features of each parameter (such as the latent heat release period of the PCM and the temperature fluctuation period caused by fluid circulation) can be fully captured by the deep neural network during the independent encoding process. This processing method not only retains the potential correlation between parameters (reconstructed through subsequent chain reasoning), but also avoids the mutual interference of different modal data at the original signal level. For example, the high-frequency noise of the ambient temperature will not directly contaminate the trend analysis of the inlet temperature.

[0034] In step S22, the PCM temperature time series, the inlet temperature time series, the outlet temperature time series, and the ambient temperature time series are encoded by time series to obtain the PCM temperature time series related implicit feature encoding vector, the inlet temperature time series related implicit feature encoding vector, the outlet temperature time series related implicit feature encoding vector, and the ambient temperature time series related implicit feature encoding vector. Specifically, in the embodiment of the present application, step S22 includes: respectively performing time series encoding based on dilated causal convolution on the PCM temperature time series, the inlet temperature time series, the outlet temperature time series, and the ambient temperature time series to obtain the PCM temperature time series related implicit feature encoding vector, the inlet temperature time series related implicit feature encoding vector, the outlet temperature time series related implicit feature encoding vector, and the ambient temperature time series related implicit feature encoding vector.

[0035] Accordingly, considering that the time series of the inlet temperature, outlet temperature, PCM temperature, and ambient temperature not only contain physical dimension differences, but also involve the coupling effect of multiple nonlinear mechanisms such as latent heat release during phase change, fluid heat transfer hysteresis, and ambient heat exchange in their dynamic evolution process. When traditional methods directly input the original time series data into the model, due to sensor noise, sampling frequency differences, and the mismatch of time scales in physical processes, it is difficult for the model to distinguish the periodic fluctuations under normal conditions from the characteristic patterns of abnormal events in the mixed original signals. For example, the slight changes in the PCM temperature during the phase change plateau period may contain early signs of material performance degradation, but these key features are easily masked by the high-frequency fluctuations of the inlet temperature or the sudden change trend of the ambient temperature. Therefore, in this application, time series encoding is performed on the PCM temperature time series, the inlet temperature time series, the outlet temperature time series, and the ambient temperature time series to obtain an inlet temperature time series correlation implicit feature encoding vector, an outlet temperature time series correlation implicit feature encoding vector, an ambient temperature time series correlation implicit feature encoding vector, and a PCM temperature time series correlation implicit feature encoding vector. In particular, in a specific example of this application, time series encoding based on dilated causal convolution is performed on the PCM temperature time series, the inlet temperature time series, the outlet temperature time series, and the ambient temperature time series respectively to obtain the PCM temperature time series correlation implicit feature encoding vector, the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector. Specifically, time series encoding is performed using time series encoding based on dilated causal convolution, and a hierarchical receptive field encoding structure is constructed by stacking convolution layers with exponentially increasing dilation rates. For each independently grouped temperature time series (such as the inlet temperature), the dilated convolution kernel samples the input values at specific intervals on the time axis, enabling the shallow network to capture local fluctuation patterns (such as temperature oscillations caused by fluid circulation), and the deep network to extract long-range trend features (such as baseline drift caused by the heat capacity decay of the phase change material). Causality is strictly ensured through the design of one-way convolution kernels, ensuring that the encoding features at each time point only depend on historical data. This design is particularly suitable for the hysteresis characteristics of the PCM temperature during the phase change stage - its temperature change is not only related to the current heat transfer state, but also affected by the cumulative effect of the previous energy storage / discharge process. Through multi-scale feature fusion, dilated causal convolution maps the original time series data into an implicit vector containing multi-dimensional information of the physical process.

[0036] The following is a detailed elaboration of a specific implementation process for "performing time series encoding based on dilated causal convolution on the PCM temperature time series, the inlet temperature time series, the outlet temperature time series, and the ambient temperature time series respectively to obtain the PCM temperature time series correlation implicit feature encoding vector, the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector":

[0037] First, data preprocessing is performed to unify the input data format. Since there are differences in the measurement ranges and precisions of different temperature sensors, the numerical distributions of each time series may have different dimensions and dynamic ranges, directly affecting the feature extraction effect of subsequent convolution operations. Therefore, normalization processing needs to be performed on the four time series respectively. The min-max normalization method is adopted. For each time series, the maximum and minimum values of its historical data are calculated, and each data point in the series is converted into a value within the range of [0, 1]. The specific operation is to subtract the minimum value of the series from each original data point and then divide by the difference between the maximum value and the minimum value of the series, thereby eliminating the influence of dimensions and making the numerical values of different series on the same scale, providing a unified input basis for subsequent convolution operations.

[0038] After completing the data normalization, enter the parameter configuration stage of the dilated causal convolution layer. Dilated causal convolution is a specially designed convolutional neural network structure, and its core lies in the causality constraint and the dilation rate mechanism. The causality constraint ensures that the output of each time point depends only on the data of the current and past time steps by setting a unidirectional convolutional kernel, avoiding the introduction of future information, which is crucial for the real-time processing of time series data. The dilation rate mechanism expands the receptive field by inserting intervals between the sampling points of the convolutional kernel, enabling the convolutional operation to capture long-range time dependencies without increasing the number of parameters. Specifically, when implementing, first determine the basic size of the convolutional kernel, usually choosing an odd length (such as 3 or 5) to ensure a symmetric context window; then set exponentially increasing dilation rates (such as 1, 2, 4, 8, etc.) for different convolutional layers to form a hierarchical multi-scale feature extraction structure. For example, the first layer of convolution uses a dilation rate of 1 to directly process the local features of adjacent time steps; the second layer uses a dilation rate of 2, expanding the effective receptive field of the convolutional kernel to 5 time steps (3×2 - 1); subsequent layers follow this pattern, and through multi-layer stacking, a comprehensive capture of short-term fluctuations and long-term trends in the time series is achieved.

[0039] Next, perform dilated causal convolution operations on the four normalized time series respectively. Taking the inlet temperature time series as an example, the configured convolution kernel starts from the starting position of the sequence and slides only in the future direction according to the causality constraint (i.e., without crossing the current time step). At each time point t, the convolution kernel covers the time steps from t-k to t (k is one less than the effective length of the convolution kernel), and performs a weighted sum of the input values at each position. The weights here are determined by the parameters of the convolution kernel, and these parameters are automatically optimized through model training to capture the key patterns in the temperature sequence. For boundary positions (such as the first k time steps at the start of the sequence), the output sequence length is kept the same as the input by padding with zero values or copying the boundary values. As the convolution kernel slides, the local context information at each time step is converted into the corresponding convolution output value, forming a preliminary feature map sequence.

[0040] After completing the linear convolution operation, a non-linear activation function needs to be applied to the feature map sequence to introduce complex feature relationships. The ReLU (Rectified Linear Unit) is selected as the activation function, and its mathematical expression is f(x) = max(0, x). This function sets the negative values in the convolution output to zero and retains the positive part, thereby enhancing the model's ability to express non-linear features without increasing the computational complexity. For example, when there is an abnormal temperature rise pulse in the inlet temperature sequence, the ReLU function can highlight this positive fluctuation feature and suppress irrelevant negative fluctuations or noise interference. The activated feature map sequence not only retains the local correlation information of the time series but also reveals the directional features of temperature changes through non-linear transformation, providing richer inputs for subsequent feature integration.

[0041] To further optimize the feature representation and reduce the data dimension, perform pooling operations on the activated feature map sequence. The max pooling method is adopted, and the maximum value is selected as the output within a preset pooling window (such as a window of size 2), thereby retaining the strongest feature responses in each local region. The pooling operation has two functions: on the one hand, it reduces the feature dimension through downsampling and reduces the model's computational load; on the other hand, it enhances the robustness of key features by ignoring secondary features. For example, in the feature map of the outlet temperature sequence, the pooling operation can filter out small fluctuations caused by high-frequency noise and retain the main trend features of temperature changes. The moving step size of the pooling window is usually the same as the window size to avoid feature overlap and information redundancy, ensuring that each element of the output sequence corresponds to a non-overlapping local region of the input sequence.

[0042] After the above convolution, activation, and pooling operations, the obtained feature map sequence has initially extracted the multi-scale temporal correlation features of the time series. However, it still needs to be converted into an encoded vector suitable for subsequent feature interaction through vectorization. The vectorization process is divided into two steps: First, flatten the feature map sequence to convert the two-dimensional sequence data (time steps × feature channels) into a one-dimensional vector. Then, adjust the dimension of the flattened vector through a fully connected layer to meet the input requirements of the subsequent model. Specifically, assume that after multiple layers of dilated causal convolution and pooling, the dimension of the feature map sequence of the inlet temperature time series is T×C (T is the number of time steps, and C is the number of feature channels). After flattening, a one-dimensional vector with a length of T×C is obtained, and then it is mapped into an implicit feature encoding vector with a preset dimension (such as a vector with a dimension of d) through the weight matrix of the fully connected layer. This mapping process can compress the global temporal correlation information of the time series into a low-dimensional vector space through training and learning, while retaining key anomaly-sensitive features.

[0043] For the PCM temperature, outlet temperature, and ambient temperature time series, the above processing flow is exactly the same: First, perform independent normalization to eliminate the influence of dimensions, and then extract their respective temporal correlation features through a dilated causal convolution network with the same structure (sharing network parameters or training independently, depending on the model design), and finally obtain their respective temporal correlation implicit feature encoding vectors.

[0044] In step S3, based on the inlet temperature temporal correlation implicit feature encoding vector, the outlet temperature temporal correlation implicit feature encoding vector, and the ambient temperature temporal correlation implicit feature encoding vector, obtain the PCM temperature simulation inference temporal correlation implicit feature. Specifically, in the embodiment of the present application, step S3 includes: performing PCM temperature simulation inference based on the inlet temperature temporal correlation implicit feature encoding vector, the outlet temperature temporal correlation implicit feature encoding vector, and the ambient temperature temporal correlation implicit feature encoding vector guided by noise-reduced inlet and outlet temperature time series to obtain the PCM temperature simulation inference temporal correlation implicit feature. More specifically, Figure 3 It is a flowchart of step S3 in the phase change energy storage thermal reservoir anomaly detection method based on deep learning according to the embodiment of the present application. As Figure 3As shown, step S3 includes: S31, respectively calculating the position-by-position difference vectors between the import temperature time-series correlation implicit feature coding vector and the export temperature time-series correlation implicit feature coding vector and the ambient temperature time-series correlation implicit feature coding vector to obtain a noise-reduced import temperature time-series correlation implicit feature coding vector and a noise-reduced export temperature time-series correlation implicit feature coding vector; S32, performing PCM temperature simulation inference based on local correlation chain on the noise-reduced import temperature time-series correlation implicit feature coding vector and the noise-reduced export temperature time-series correlation implicit feature coding vector to obtain a PCM temperature simulation inference time-series correlation implicit feature coding vector as the PCM temperature simulation inference time-series correlation implicit feature.

[0045] In step S31, the position-wise difference vectors between the imported temperature time-series associated implicit feature encoding vector and the exported temperature time-series associated implicit feature encoding vector and the ambient temperature time-series associated implicit feature encoding vector are calculated respectively to obtain a noise-reduced imported temperature time-series associated implicit feature encoding vector and a noise-reduced exported temperature time-series associated implicit feature encoding vector. Correspondingly, considering the ambient temperature as an external disturbance source, its dynamic fluctuations will be coupled into the signals of the imported temperature and the exported temperature through physical mechanisms such as heat conduction and convection, forming noises mixed with the true abnormal features inside the system. For example, the change in the day-night temperature difference may cause fluctuations in the heat exchange intensity between the heat transfer fluid and the environment, resulting in periodic disturbances in the imported / exported temperature that are unrelated to equipment aging. Due to the time-delay effect and non-linear response in the fluid heat transfer process, simple linear subtraction or filtering operations cannot accurately distinguish the actual impact of the ambient temperature on the system's heat transfer link, and may instead damage the dynamic correlation between temperature parameters. Therefore, in the technical solution of this application, the position-wise difference vectors between the imported temperature time-series associated implicit feature encoding vector and the exported temperature time-series associated implicit feature encoding vector and the ambient temperature time-series associated implicit feature encoding vector are calculated respectively to obtain a noise-reduced imported temperature time-series associated implicit feature encoding vector and a noise-reduced exported temperature time-series associated implicit feature encoding vector. In particular, after the independent time-series encoding of each temperature parameter, the feature vectors of the imported temperature and the exported temperature already contain the thermodynamic correlation features of their own time-series evolution laws, while the ambient temperature feature vector encodes the propagation mode of external disturbances. The essence of the position-wise difference operation is to deduct the influence weight of the ambient temperature feature on the heat transfer link dimension by dimension from the imported / exported temperature feature space. This operation is not a simple signal subtraction, but rather, through the implicit mapping relationship established by the deep learning model during the encoding stage, it identifies the coupling components of the ambient temperature fluctuations in the imported / exported temperature features, such as the delayed interference features caused by the sudden drop in the ambient temperature through the conduction of the heat reservoir shell to the fluid inlet temperature. In this way, the obtained noise-reduced imported temperature time-series associated implicit feature encoding vector and noise-reduced exported temperature time-series associated implicit feature encoding vector remove the noises brought by the ambient temperature interference, and the feature quality is significantly improved. These cleaner feature vectors can more accurately reflect the heat exchange process and operating state inside the heat reservoir, providing a more reliable data basis for subsequent analysis and modeling.

[0046] In step S32, a PCM temperature simulation inference based on local association chaining is performed on the noise-reduced imported temperature time-series associated implicit feature encoding vector and the noise-reduced exported temperature time-series associated implicit feature encoding vector to obtain a PCM temperature simulation inference time-series associated implicit feature encoding vector as the PCM temperature simulation inference time-series associated implicit feature. Specifically, Figure 4 FIG. is a flowchart of step S32 in the method for detecting anomalies in a phase change energy storage heat reservoir based on deep learning according to an embodiment of the present application. AsFigure 4 As shown in Figure 4 , step S32 includes: S32-1, respectively performing local implicit feature extraction based on one-dimensional convolutional encoding on the noise-reduced inlet temperature time-series correlation implicit feature encoding vector and the noise-reduced outlet temperature time-series correlation implicit feature encoding vector to obtain a set of noise-reduced inlet temperature local time-series implicit feature encoding vectors and a set of noise-reduced outlet temperature local time-series implicit feature encoding vectors; S32-2, performing chain inference attention weight modulation based on feature interaction response on the set of noise-reduced inlet temperature local time-series implicit feature encoding vectors and the set of noise-reduced outlet temperature local time-series implicit feature encoding vectors to obtain a set of PCM temperature simulation modulation local time-series implicit feature interaction response encoding vectors; S32-3, inputting the set of PCM temperature simulation modulation local time-series implicit feature interaction response encoding vectors into a chain inference engine based on a forward LSTM model to obtain the PCM temperature simulation inference time-series correlation implicit feature encoding vector.

[0047] It should be understood that in the phase change energy storage heat reservoir, there are complex thermodynamic relationships among the inlet temperature, the outlet temperature, and the PCM temperature. These relationships are not simple linear superpositions, but rather have complex dynamic correlations such as time-varying coupling, thermodynamic hysteresis, and differences in energy transfer paths. When traditional methods use linear regression or fixed transfer function modeling, they cannot capture the non-linear perturbation of the sudden change in the inlet temperature on the PCM phase change plateau period, nor can they quantify the energy release hysteresis effect between the outlet temperature and the PCM temperature. For example, the deterioration of the microstructure of the phase change material will cause a decrease in its latent heat release rate, but this anomaly only manifests as a slight deviation in the dynamic response between the inlet-outlet temperature and the PCM temperature in the early stage, and single-parameter threshold monitoring or simple correlation analysis is extremely likely to miss such hidden faults. Therefore, in order to more accurately predict the PCM temperature characteristics based on the inlet and outlet temperature characteristics, this application performs PCM temperature simulation inference based on local association chain on the noise-reduced inlet temperature time-series correlation implicit feature encoding vector and the noise-reduced outlet temperature time-series correlation implicit feature encoding vector to obtain the PCM temperature simulation inference time-series correlation implicit feature encoding vector as the PCM temperature simulation inference time-series correlation implicit feature.

[0048] Specifically, first, one-dimensional convolution is used to extract the local patterns of the denoised inlet temperature time-series associated implicit feature encoding vector and the denoised outlet temperature time-series associated implicit feature encoding vector (such as the temperature oscillation period caused by fluid circulation and the transient heat transfer pulse characteristics). Subsequently, the dynamic coupling relationship between the two within the local time window is modeled through the single-entity feature interaction engine (such as the multiplicative interaction of the inlet temperature pulse on the outlet temperature hysteresis effect). The chain inference attention mechanism further screens the key interaction nodes (such as the feature segments corresponding to the phase change plateau period), and integrates the interaction effects across time scales through the time-series memory ability of the forward LSTM (such as the long-range effect of the inlet temperature cumulative effect on the PCM phase change process). This process essentially establishes an implicit model of the physical constraint relationship between the inlet-outlet temperature and the PCM temperature under normal conditions, transforming the thermodynamic constitutive equation into a learnable feature space mapping rule.

[0049] Specifically, in the embodiment of the present application, the step S32-1 includes: respectively performing local implicit feature extraction based on one-dimensional convolution encoding on the denoised inlet temperature time-series associated implicit feature encoding vector and the denoised outlet temperature time-series associated implicit feature encoding vector to obtain a set of denoised inlet temperature local time-series implicit feature encoding vectors and a set of denoised outlet temperature local time-series implicit feature encoding vectors, which can be expressed by the following formula:

[0050] Conv l×1 (X) = {x1, x2,..., x i ,..., x n}

[0051] Conv l×1 (Y) = {y1, y2,..., y i ,..., y n}

[0052] Wherein, X is the denoised inlet temperature time-series associated implicit feature encoding vector, Y is the denoised outlet temperature time-series associated implicit feature encoding vector, Conv l×1 is the local implicit feature extraction based on one-dimensional convolution, l is the length of the one-dimensional convolution kernel, x1, x2, x i and x n are respectively the 1st, 2nd, ith, and nth denoised inlet temperature local time-series implicit feature encoding vectors in the set of denoised inlet temperature local time-series implicit feature encoding vectors, y1, y2, y i and y n are respectively the 1st, 2nd, ith, and nth denoised outlet temperature local time-series implicit feature encoding vectors in the set of denoised outlet temperature local time-series implicit feature encoding vectors, n is the number of vectors in X and Y, and X and Y have the same length.

[0053] It should be understood that in the phase change energy storage thermal reservoir, the dynamic inlet and outlet temperatures contain local time-varying patterns (such as temperature oscillation periods, transient heat transfer pulses), which are closely related to the non-linear response of the PCM phase change process (such as changes in the latent heat release rate). Traditional linear methods cannot capture such local features, while one-dimensional convolution, through a sliding window mechanism with parameter sharing, can efficiently extract the local correlation features hidden in the temperature sequence (such as periodic fluctuations caused by fluid circulation or transient effects of sudden changes in the inlet temperature). That is, by designing convolution kernels of different sizes, the model can not only capture fine-grained microscopic temperature fluctuations (such as early signals of the decrease in the latent heat release rate caused by the deterioration of the PCM microstructure), but also cover the macroscopic heat transfer trend (such as the overall temperature distribution during the phase change plateau period). In this way, the set of denoised local time series implicit feature encoding vectors of the inlet temperature and the set of denoised local time series implicit feature encoding vectors of the outlet temperature can provide denoised local time series features for subsequent interaction modeling, avoiding interference from the original noise on high-order interaction reasoning.

[0054] Specifically, Figure 5 It is a flowchart of step S32-2 in the phase change energy storage thermal reservoir anomaly detection method based on deep learning according to an embodiment of the present application. As Figure 5 shown, the step S32-2 includes: S32-21, respectively inputting each group of corresponding denoised local time series implicit feature encoding vectors of the inlet temperature and the denoised local time series implicit feature encoding vectors of the outlet temperature in the set of denoised local time series implicit feature encoding vectors of the inlet temperature and the set of denoised local time series implicit feature encoding vectors of the outlet temperature into a single feature interaction engine to obtain a set of denoised inlet-outlet temperature local time series implicit feature interaction response encoding vectors; S32-22, based on the feature distribution characteristics of each denoised inlet-outlet temperature local time series implicit feature interaction response encoding vector in the set of denoised inlet-outlet temperature local time series implicit feature interaction response encoding vectors, determining the chain reasoning attention weights of each denoised inlet-outlet temperature local time series implicit feature interaction response encoding vector to obtain a set of PCM temperature simulation chain reasoning attention weights; S32-23, based on the set of PCM temperature simulation chain reasoning attention weights, performing weighted modulation on the set of denoised inlet-outlet temperature local time series implicit feature interaction response encoding vectors to obtain a set of PCM temperature simulation modulated local time series implicit feature interaction response encoding vectors.

[0055] More specifically, in the embodiments of the present application, the step S32-21 includes: respectively inputting each corresponding pair of the denoised inlet temperature local temporal implicit feature encoding vectors and the denoised outlet temperature local temporal implicit feature encoding vectors in the set of the denoised inlet temperature local temporal implicit feature encoding vectors and the set of the denoised outlet temperature local temporal implicit feature encoding vectors into the single feature interaction engine to obtain a set of denoised inlet-outlet temperature local temporal implicit feature interaction response encoding vectors, which can be expressed by the following formula:

[0056]

[0057] wherein, ⊙ is element-wise multiplication, is element-wise addition, is element-wise subtraction, concat{·;·;·} is a concatenation operation, and W i is the i-th denoised inlet-outlet temperature implicit interaction response weight matrix in the set of denoised inlet-outlet temperature implicit interaction response weight matrices, and b i is the i-th denoised inlet-outlet temperature implicit interaction response bias vector in the set of denoised inlet-outlet temperature implicit interaction response bias vectors, and v i is the i-th denoised inlet-outlet temperature local temporal implicit feature interaction response encoding vector in the set of denoised inlet-outlet temperature local temporal implicit feature interaction response encoding vectors.

[0058] It should be understood that there is a time-varying coupling in the thermodynamic relationship between the inlet temperature and the outlet temperature (for example, the inlet temperature pulse affects the outlet temperature through the phase change lag of the PCM). Through element-wise interaction, the single feature interaction engine can explicitly model the dynamic relationship between the two feature vectors within a local time window, such as simulating the multiplicative modulation of the hysteresis effect of the inlet temperature mutation on the outlet temperature. That is, this step transforms the physical constraints (such as energy conservation and heat transfer path differences) implicit in the thermodynamic constitutive equation into interaction rules in the feature space, enhancing the model's ability to express dynamic coupling effects. At the same time, by performing interaction at the local implicit feature layer, redundant noise (such as sensor jitter) in the original features can be filtered, focusing on the correlation of the abstracted heat transfer patterns.

[0059] More specifically, in the embodiments of the present application, the step S32-22 includes: based on the feature distribution characteristics of each denoised inlet-outlet temperature local temporal implicit feature interaction response encoding vector in the set of denoised inlet-outlet temperature local temporal implicit feature interaction response encoding vectors, determining the chain inference attention weights of each denoised inlet-outlet temperature local temporal implicit feature interaction response encoding vector to obtain a set of PCM temperature simulation chain inference attention weights, which can be expressed by the following formula:

[0060]

[0061] wherein, v i,j is the j-th eigenvalue in v i , ||·|| 2 is the square of the Euclidean norm for calculating a vector, L is the number of eigenvalues in v i , Softmax is the Softmax function, a i is the i-th PCM temperature simulation chain inference attention weight in the set of PCM temperature simulation chain inference attention weights.

[0062] It should be understood that the local interaction importance in different time periods during the phase change process varies significantly (for example, the characteristic segments during the phase change plateau are more sensitive to faults). Calculating the PCM temperature simulation chain inference attention weights based on the characteristic distribution characteristics can quantify the information density and abnormal contribution degree of the local time series implicit feature interaction response nodes of the noise-reduced inlet-outlet temperature. For example, the abnormal latent heat release caused by the deterioration of the PCM microstructure may be manifested as the distribution shift of the interaction response vector within a specific time window. That is, by dynamically allocating weights, the key nodes (such as the interaction features corresponding to the phase change plateau) are highlighted, and the redundant information in the steady-state heat transfer stage is weakened, enabling the model to focus on the weak signals of implicit faults and improving the sensitivity to early latent abnormalities (such as the decrease in the latent heat release rate).

[0063] More specifically, in the embodiment of the present application, the step S32-23 includes: performing multi-modal attention correction of the phase change process based on interaction norm decomposition on the set of PCM temperature simulation chain inference attention weights to obtain a set of corrected PCM temperature simulation chain inference attention weights, and this process can be expressed by the formula:

[0064]

[0065] a′ i =(ω1×ξ i +ω2×ζ i )a i

[0066] wherein, α i is the i-th PCM temperature simulation chain inference translational action interaction energy intensity in the set of PCM temperature simulation chain inference translational action interaction energy intensities, β i is the i-th PCM temperature simulation chain inference translational action difference energy intensity in the set of PCM temperature simulation chain inference translational action difference energy intensities, γ i is the i-th PCM temperature simulation chain inference fluctuation action energy intensity in the set of PCM temperature simulation chain inference fluctuation action energy intensities, ln is the logarithmic function value with the natural constant e as the base, ξ iis the i-th PCM temperature simulation chain inference periodic local rule compensation factor in the set of PCM temperature simulation chain inference periodic local rule compensation factors, ζ i is the i-th PCM temperature simulation chain inference translational periodic local auxiliary phase change factor in the set of PCM temperature simulation chain inference translational periodic local auxiliary phase change factors, ω1 and ω2 are the compensation dynamic weight coefficient and the phase change dynamic weight coefficient respectively, a′ i is the i-th corrected PCM temperature simulation chain inference attention weight in the set of corrected PCM temperature simulation chain inference attention weights;

[0067] Taking the set of the corrected PCM temperature simulation chain inference attention weights as the set of weights, weighting and modulating the set of the denoised inlet-outlet temperature local temporal implicit feature interaction response encoding vectors to obtain the set of the PCM temperature simulation modulated local temporal implicit feature interaction response encoding vectors, this process can be expressed by the formula:

[0068] I i = a′ i ·v i

[0069] I = {I1, I2,..., I i ,..., I n}

[0070] wherein, I1, I2, I i and I n are the 1st, 2nd, i-th and n-th PCM temperature simulation modulated local temporal implicit feature interaction response encoding vectors in the set of PCM temperature simulation modulated local temporal implicit feature interaction response encoding vectors respectively, and I is the set of PCM temperature simulation modulated local temporal implicit feature interaction response encoding vectors.

[0071] Particularly, when calculating each denoised inlet-outlet temperature local temporal implicit feature interaction response encoding vector, the interaction feature between the corresponding denoised inlet temperature local temporal implicit feature encoding vector and the denoised outlet temperature local temporal implicit feature encoding vector will be introduced, such as x i ⊙y i , etc., that is to say, it actually corresponds to different interaction specifications, so as to have different spatial auxiliary constraint correlations in the interaction space.

[0072] Therefore, in order to enhance the specification variability assistance when inferring the attention weights in the chain, preferably based on the spatial effect decomposition of the interaction specification, the PCM temperature simulation chain inference attention weights are corrected. Specifically, It is regarded as the translational effect of the PCM temperature simulation chain inference, that is, the gradient direction is consistent with the spatial interaction direction, and x i ⊙y i is regarded as the fluctuation effect of the PCM temperature simulation chain inference, that is, the gradient direction is orthogonal to the spatial interaction direction. In this way, for the vector statistics of different PCM temperature simulation chain inference effects, such as γ i =||x i ⊙y i || 2 , it is considered that the fluctuation effect of the PCM temperature simulation chain inference causes a localized periodic effect in the translational direction, and thus calculates its PCM temperature simulation chain inference periodic local rule compensation factor as:

[0073]

[0074] That is, α + β is used as the translational localization representation, and the fluctuation effect of the PCM temperature simulation chain inference is enhanced as the logarithmic size of the translational localization increases.

[0075] On the other hand, the fluctuation effect of the PCM temperature simulation chain inference will also cause a phase change in the translational localization. Therefore, the PCM temperature simulation chain inference translational periodic local auxiliary phase change factor is obtained as:

[0076]

[0077] Thus, based on the weighted sum of the above two items, the attention weight a of the PCM temperature simulation chain inference is corrected i :

[0078] a′ i =(ω1×ξ i +ω2×ζ i )a i

[0079] It is possible to improve the correlation between different spatial auxiliary constraint specifications in the interaction space according to the difference in the normative effects of different interaction specifications corresponding to each local interaction response in the interaction space, thereby improving the calculation accuracy of the attention weight of the PCM temperature simulation chain inference.

[0080] It should be understood that the accuracy of PCM temperature prediction depends on the precise capture of key interaction nodes. By weighting and modulating the denoised inlet - outlet temperature local temporal implicit feature interaction response encoding vector with the corrected PCM temperature simulation chain - inference attention weights, the contribution of high - weight vectors (such as interaction features during abnormal periods) can be amplified, and low - weight noise (such as regular fluctuations in steady - state heat transfer) can be suppressed. For example, the corrected weight set can strengthen the interaction pattern shift caused by PCM structure deterioration (such as the asymmetric change in the inlet - outlet temperature response within a certain time window), making the modulated feature vector more significantly represent the abnormal state. That is to say, this processing process essentially maps the fault sensitivity of the physical system to the feature space and improves the model's ability to represent latent faults through an information - selection mechanism.

[0081] Specifically, in the embodiment of the present application, in step S32 - 3, the set of PCM temperature simulation - modulated local temporal implicit feature interaction response encoding vectors is input into a chain - inference engine based on a forward LSTM model to obtain the PCM temperature simulation - inference temporal - correlation implicit feature encoding vector, which can be expressed by the following formula:

[0082]

[0083] Where, is the forward LSTM encoding, and v f is the PCM temperature simulation - inference temporal - correlation implicit feature encoding vector.

[0084] It should be understood that the PCM phase - change process has long - range temporal dependence (such as the cumulative effect of inlet temperature affecting the subsequent phase - change plateau period). The forward LSTM can gradually integrate the set of PCM temperature simulation - modulated local temporal implicit feature interaction response encoding vectors through a gating mechanism (forget gate, input gate) to capture the thermodynamic laws across time scales. For example, the memory unit of LSTM can store the short - term impact effect of the inlet - temperature pulse and at the same time associate it with the long - term correlation of the subsequent PCM temperature hysteresis release. That is to say, this step converts the discrete features of local interactions (such as transient heat - transfer pulses) into globally coherent temporal - inference results, simulates the energy - transfer chain reaction of the heat - storage system, and the finally output PCM temperature simulation - inference temporal - correlation implicit feature encoding vector can represent the comprehensive state of the PCM temperature dynamic evolution (including normal phase - change and early abnormal deviation).

[0085] In step S4, based on the PCM temperature simulation - inference temporal - correlation implicit feature and the PCM temperature temporal - correlation implicit feature encoding vector, it is determined whether there is an abnormality in the phase - change energy - storage heat reservoir. Specifically, Figure 6 is the flowchart of step S4 in the method for abnormal detection of a phase - change energy - storage heat reservoir based on deep learning according to the embodiment of the present application. As Figure 6As shown, step S4 includes: S41, calculating the position-wise difference vector between the PCM temperature simulation inference time-series associated implicit feature encoding vector and the PCM temperature time-series associated implicit feature encoding vector to obtain the PCM temperature inference simulation-real difference encoding vector; S42, inputting the PCM temperature inference simulation-real difference encoding vector into an anomaly analyzer based on a classifier to determine whether there is an anomaly in the phase change energy storage heat reservoir.

[0086] In step S41, the position-wise difference vector between the PCM temperature simulation inference time-series associated implicit feature encoding vector and the PCM temperature time-series associated implicit feature encoding vector is calculated to obtain the PCM temperature inference simulation-real difference encoding vector. Correspondingly, considering that the PCM temperature simulation inference time-series associated implicit feature encoding vector is the feature that the PCM temperature should have under normal working conditions simulated based on the denoised inlet and outlet temperature features. While the PCM temperature time-series associated implicit feature encoding vector is the real feature extracted from the actually monitored PCM temperature time series. The anomalies in the thermodynamic system often manifest as latent deviations between the actual operating state and the expected behavior under normal working conditions. For example, the deterioration of the microstructure of the phase change material will change its latent heat release characteristics, but such anomalies may only cause slight offsets in the dynamic response between the PCM temperature and the inlet / outlet temperature in the early stage, and it is difficult to capture such complex faults by the absolute threshold or trend analysis of a single temperature parameter. Based on this, in the technical solution of this application, the position-wise difference vector between the PCM temperature simulation inference time-series associated implicit feature encoding vector and the PCM temperature time-series associated implicit feature encoding vector is calculated to obtain the PCM temperature inference simulation-real difference encoding vector. In this way, the difference between the model simulation result and the actual situation can be intuitively compared, and the obtained PCM temperature inference simulation-real difference encoding vector provides a clear quantitative index for anomaly detection. By setting an appropriate threshold, it can be determined whether the difference vector exceeds the normal range. If it exceeds the threshold, it indicates that there may be an abnormal situation in the heat reservoir. This judgment method based on quantitative differences improves the accuracy and reliability of anomaly detection.

[0087] In step S42, the PCM temperature inference simulation - real difference coding vector is input into the classifier - based anomaly analyzer to determine whether there is an anomaly in the phase - change energy - storage thermal reservoir. It should be understood that a classifier is a tool widely used in the field of machine learning, which can classify input data into different categories according to the characteristic patterns of the input data. The PCM temperature inference simulation - real difference coding vector contains key information about the operating state of the thermal reservoir. The classifier can learn these complex characteristic relationships and discover the patterns hidden in the data. For example, certain combinations of feature dimensions may be closely related to specific types of anomalies. The classifier can capture these combined patterns through training, thereby making more accurate anomaly judgments. In this way, once the classifier determines that there is an anomaly in the thermal reservoir, a warning signal can be sent in a timely manner to notify the operator to take corresponding measures. This helps to detect and handle problems in the operation of the thermal reservoir in a timely manner, avoid the further deterioration of problems, reduce losses, and ensure the safe and stable operation of the thermal reservoir. Specifically, in a specific embodiment of the present application, inputting the PCM temperature inference simulation - real difference coding vector into the classifier - based anomaly analyzer to determine whether there is an anomaly in the phase - change energy - storage thermal reservoir includes: performing fully - connected coding on the PCM temperature inference simulation - real difference coding vector using the fully - connected layer of the classifier to obtain the PCM temperature inference simulation - real difference fully - connected coding feature vector; inputting the PCM temperature inference simulation - real difference fully - connected coding feature vector into the Softmax classification function of the classifier to obtain the probability values of the PCM temperature inference simulation - real difference coding vector belonging to each classification label, where the classification labels include those for indicating the existence of an anomaly in the phase - change energy - storage thermal reservoir and those for indicating the non - existence of an anomaly in the phase - change energy - storage thermal reservoir; and determining the classification label corresponding to the largest probability value as the anomaly - detection classification result.

[0088] In summary, the deep - learning - based phase - change energy - storage thermal reservoir anomaly - detection method according to the embodiments of the present application is elucidated. First, independent temporal encoding is performed on the PCM temperature, inlet and outlet temperatures, and ambient temperature to extract the time - correlation features of each parameter. Then, bit - by - bit difference operation is performed on the ambient - temperature temporal - correlation hidden feature and the inlet - and - outlet - temperature temporal - correlation hidden feature to strip the noise interference of ambient - temperature fluctuations on the heat - transfer link from the feature space. Next, a thermodynamic response model between the inlet - outlet temperature and the PCM temperature is constructed using chained - inference coding to generate the PCM temperature simulation features based on normal operating conditions. Finally, by comparing the difference vector between the simulation features and the actual features, the deviation degree detection of the thermodynamic constitutive relationship after ambient - noise decoupling is realized. In this way, the detection sensitivity of anomalies in the phase - change energy - storage thermal reservoir can be significantly improved.

[0089] Figure 7 FIG. is a system block diagram of a deep - learning - based phase - change energy - storage thermal reservoir anomaly - detection system according to an embodiment of the present application. As Figure 7As shown in the figure, the abnormal detection system 100 of a phase change energy storage thermal reservoir based on deep learning according to an embodiment of the present application includes: a running data extraction module 110, configured to extract a time series data set of running data from a monitoring system of the phase change energy storage thermal reservoir, where the running data includes PCM temperature, inlet temperature, outlet temperature, and ambient temperature; a data encoding and grouping module 120, configured to perform data grouping and time series encoding on the time series data set of the running data to obtain an inlet temperature time series associated hidden feature encoding vector, an outlet temperature time series associated hidden feature encoding vector, an ambient temperature time series associated hidden feature encoding vector, and a PCM temperature time series associated hidden feature encoding vector; a PCM temperature simulation and inference module 130, configured to obtain a PCM temperature simulation inference time series associated hidden feature based on the inlet temperature time series associated hidden feature encoding vector, the outlet temperature time series associated hidden feature encoding vector, and the ambient temperature time series associated hidden feature encoding vector; and an abnormality determination module 140, configured to determine whether there is an abnormality in the phase change energy storage thermal reservoir based on the PCM temperature simulation inference time series associated hidden feature and the PCM temperature time series associated hidden feature encoding vector.

[0090] Here, those skilled in the art can understand that the specific functions and operations of each unit and module in the above-mentioned abnormal detection system 100 of a phase change energy storage thermal reservoir based on deep learning have been introduced in detail in the description of the Figures 1 to 6 abnormal detection method of a phase change energy storage thermal reservoir based on deep learning, and therefore, the repeated description thereof will be omitted.

[0091] In summary, the abnormal detection system 100 of a phase change energy storage thermal reservoir based on deep learning according to an embodiment of the present application is clarified. First, independent time series encoding is performed on the PCM temperature, inlet and outlet temperatures, and ambient temperature to extract the time correlation features of each parameter. Then, bitwise difference operation is performed on the ambient temperature time series associated hidden feature and the inlet and outlet temperature time series associated hidden features to strip the noise interference of the ambient temperature fluctuation on the heat transfer link from the feature space. Next, a thermodynamic response model between the inlet-outlet temperature and the PCM temperature is constructed by using chain inference encoding to generate the PCM temperature simulation features based on the normal working condition. Finally, by comparing the difference vector between the simulation features and the actual features, the deviation degree detection of the thermodynamic constitutive relationship after environmental noise decoupling is realized. In this way, the detection sensitivity of the abnormality of the phase change energy storage thermal reservoir can be significantly improved.

Claims

1. A method for detecting anomalies in a phase change energy storage thermal reservoir based on deep learning, characterized in that, Including: A time series dataset extracting operation data from a monitoring system of a phase change energy storage heat reservoir, where the operation data includes PCM temperature, inlet temperature, outlet temperature, and ambient temperature; Grouping and time series encoding the time series dataset of the operation data to obtain an inlet temperature time series correlation implicit feature encoding vector, an outlet temperature time series correlation implicit feature encoding vector, an ambient temperature time series correlation implicit feature encoding vector, and a PCM temperature time series correlation implicit feature encoding vector; Based on the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector, obtaining a PCM temperature simulation inference time series correlation implicit feature, including: performing PCM temperature simulation inference based on noise-reduced inlet and outlet temperature time series guidance on the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector to obtain the PCM temperature simulation inference time series correlation implicit feature; Based on the PCM temperature simulation inference time series correlation implicit feature and the PCM temperature time series correlation implicit feature encoding vector, determining whether there is an abnormality in the phase change energy storage heat reservoir.

2. The method for abnormal detection of the phase change energy storage heat reservoir based on deep learning according to claim 1, wherein Grouping and time series encoding the time series dataset of the operation data to obtain an inlet temperature time series correlation implicit feature encoding vector, an outlet temperature time series correlation implicit feature encoding vector, an ambient temperature time series correlation implicit feature encoding vector, and a PCM temperature time series correlation implicit feature encoding vector, including: Grouping the time series dataset of the operation data to obtain a PCM temperature time series, an inlet temperature time series, an outlet temperature time series, and an ambient temperature time series; Performing time series encoding on the PCM temperature time series, the inlet temperature time series, the outlet temperature time series, and the ambient temperature time series to obtain the PCM temperature time series correlation implicit feature encoding vector, the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector.

3. The method for abnormal detection of the phase change energy storage thermal reservoir based on deep learning according to claim 2, wherein Performing time series encoding on the PCM temperature time series, the inlet temperature time series, the outlet temperature time series, and the ambient temperature time series to obtain the PCM temperature time series correlation implicit feature encoding vector, the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector, including: respectively performing time series encoding based on dilated causal convolution on the PCM temperature time series, the inlet temperature time series, the outlet temperature time series, and the ambient temperature time series to obtain the PCM temperature time series correlation implicit feature encoding vector, the inlet temperature time series correlation implicit feature encoding vector, the outlet temperature time series correlation implicit feature encoding vector, and the ambient temperature time series correlation implicit feature encoding vector.

4. The method for abnormal detection of a phase change energy storage thermal reservoir based on deep learning according to claim 1, wherein Perform PCM temperature simulation inference based on noise-reduced import and export temperature time series guidance on the import temperature time series associated implicit feature encoding vector, the export temperature time series associated implicit feature encoding vector, and the ambient temperature time series associated implicit feature encoding vector to obtain the PCM temperature simulation inference time series associated implicit feature, including: Calculate the position-wise difference vectors between the import temperature time series associated implicit feature encoding vector and the export temperature time series associated implicit feature encoding vector and the ambient temperature time series associated implicit feature encoding vector respectively to obtain a noise-reduced import temperature time series associated implicit feature encoding vector and a noise-reduced export temperature time series associated implicit feature encoding vector; Perform PCM temperature simulation inference based on local association chaining on the noise-reduced import temperature time series associated implicit feature encoding vector and the noise-reduced export temperature time series associated implicit feature encoding vector to obtain a PCM temperature simulation inference time series associated implicit feature encoding vector as the PCM temperature simulation inference time series associated implicit feature.

5. The method for abnormal detection of the phase change energy storage heat reservoir based on deep learning according to claim 4, wherein Perform PCM temperature simulation inference based on local association chaining on the noise-reduced import temperature time series associated implicit feature encoding vector and the noise-reduced export temperature time series associated implicit feature encoding vector to obtain a PCM temperature simulation inference time series associated implicit feature encoding vector, including: Perform local implicit feature extraction based on one-dimensional convolutional encoding on the noise-reduced import temperature time series associated implicit feature encoding vector and the noise-reduced export temperature time series associated implicit feature encoding vector respectively to obtain a set of noise-reduced import temperature local time series implicit feature encoding vectors and a set of noise-reduced export temperature local time series implicit feature encoding vectors; Perform chained inference attention weight modulation based on feature interaction response on the set of noise-reduced import temperature local time series implicit feature encoding vectors and the set of noise-reduced export temperature local time series implicit feature encoding vectors to obtain a set of PCM temperature simulation modulated local time series implicit feature interaction response encoding vectors; Input the set of PCM temperature simulation modulated local time series implicit feature interaction response encoding vectors into a chained inference engine based on a forward LSTM model to obtain the PCM temperature simulation inference time series associated implicit feature encoding vector.

6. The method for abnormal detection of the phase change energy storage thermal reservoir based on deep learning according to claim 5, wherein, Perform chained inference attention weight modulation based on feature interaction response on the set of noise-reduced import temperature local time series implicit feature encoding vectors and the set of noise-reduced export temperature local time series implicit feature encoding vectors to obtain a set of PCM temperature simulation modulated local time series implicit feature interaction response encoding vectors, including: Input each group of corresponding noise-reduced import temperature local time series implicit feature encoding vectors and noise-reduced export temperature local time series implicit feature encoding vectors in the set of noise-reduced import temperature local time series implicit feature encoding vectors and the set of noise-reduced export temperature local time series implicit feature encoding vectors into a single feature interaction engine respectively to obtain a set of noise-reduced import-export temperature local time series implicit feature interaction response encoding vectors; Based on the feature distribution characteristics of each denoised inlet-outlet temperature local temporal implicit feature interaction response coding vector in the set of denoised inlet-outlet temperature local temporal implicit feature interaction response coding vectors, determine the chain inference attention weights of each denoised inlet-outlet temperature local temporal implicit feature interaction response coding vector to obtain a set of PCM temperature simulation chain inference attention weights; Based on the set of PCM temperature simulation chain inference attention weights, perform weighted modulation on the set of denoised inlet-outlet temperature local temporal implicit feature interaction response coding vectors to obtain the set of PCM temperature simulation modulated local temporal implicit feature interaction response coding vectors.

7. The method for abnormal detection of a phase change energy storage thermal reservoir based on deep learning according to claim 6, wherein Based on the set of PCM temperature simulation chain inference attention weights, performing weighted modulation on the set of denoised inlet-outlet temperature local temporal implicit feature interaction response coding vectors to obtain the set of PCM temperature simulation modulated local temporal implicit feature interaction response coding vectors includes: Perform multi-modal attention correction for the phase change process based on interaction specification decomposition on the set of PCM temperature simulation chain inference attention weights to obtain a set of corrected PCM temperature simulation chain inference attention weights; Using the set of corrected PCM temperature simulation chain inference attention weights as the set of weights, perform weighted modulation on the set of denoised inlet-outlet temperature local temporal implicit feature interaction response coding vectors to obtain the set of PCM temperature simulation modulated local temporal implicit feature interaction response coding vectors.

8. The method for abnormal detection of the phase change energy storage thermal reservoir based on deep learning according to claim 7, characterized in that Based on the PCM temperature simulation inference temporal correlation implicit feature and the PCM temperature temporal correlation implicit feature coding vector, determine whether there is an abnormality in the phase change energy storage heat reservoir, including: Calculate the position-wise difference vector between the PCM temperature simulation inference temporal correlation implicit feature coding vector and the PCM temperature temporal correlation implicit feature coding vector to obtain the PCM temperature inference simulation-real difference coding vector; Input the PCM temperature inference simulation-real difference coding vector into an anomaly analyzer based on a classifier to determine whether there is an abnormality in the phase change energy storage heat reservoir.

9. An abnormal detection system for a phase change energy storage thermal reservoir based on deep learning, characterized in that, Including: An operation data extraction module for extracting a temporal data set of operation data from the monitoring system of the phase change energy storage heat reservoir, where the operation data includes PCM temperature, inlet temperature, outlet temperature, and ambient temperature; A data coding and grouping module for performing data grouping and time series coding on the temporal data set of the operation data to obtain an inlet temperature temporal correlation implicit feature coding vector, an outlet temperature temporal correlation implicit feature coding vector, an ambient temperature temporal correlation implicit feature coding vector, and a PCM temperature temporal correlation implicit feature coding vector; The PCM temperature simulation inference module is used to obtain the PCM temperature simulation inference time-series associated hidden features based on the inlet temperature time-series associated hidden feature encoding vector, the outlet temperature time-series associated hidden feature encoding vector, and the ambient temperature time-series associated hidden feature encoding vector. Among them, the PCM temperature simulation inference module is used to: perform PCM temperature simulation inference based on the denoised inlet and outlet temperature time-series guidance on the inlet temperature time-series associated hidden feature encoding vector, the outlet temperature time-series associated hidden feature encoding vector, and the ambient temperature time-series associated hidden feature encoding vector to obtain the PCM temperature simulation inference time-series associated hidden features; The anomaly determination module is used to determine whether there is an anomaly in the phase change energy storage heat reservoir based on the PCM temperature simulation inference time-series associated hidden features and the PCM temperature time-series associated hidden feature encoding vector.

10. The abnormal detection system for the phase change energy storage thermal reservoir based on deep learning according to claim 9, wherein, The anomaly determination module includes: The position-wise difference calculation unit is used to calculate the position-wise difference vector between the PCM temperature simulation inference time-series associated hidden feature encoding vector and the PCM temperature time-series associated hidden feature encoding vector to obtain the PCM temperature inference simulation-real difference encoding vector; The anomaly result generation unit is used to input the PCM temperature inference simulation-real difference encoding vector into the anomaly analyzer based on the classifier to determine whether there is an anomaly in the phase change energy storage heat reservoir.

Images