A method for energy recovery of power equipment based on multimodal processing

Abstract

The present invention discloses a method for energy recovery from electric power equipment based on multimodal processing, relating to the field of electric energy data technology. The method comprises: step S1: collecting load data, electromagnetic power values, and recovered heat energy values ​​generated by target electric power equipment; step S2: screening out abnormal cycles with electromagnetic anomalies; step S3: constructing a multimodal learning network to compensate and correct the recovered heat energy values ​​within the abnormal cycles; step S4: setting a first reference heat energy value and marking the corrected heat energy value; and step S5: determining whether the recovered heat energy value within the current abnormal cycle has been corrected, if not, re-executing step S3. Compared with the prior art, the present invention constructs a multimodal learning network based on electromagnetic power values ​​and load data, which can accurately capture the energy recovery effect of the equipment under different load conditions and automatically re-correct it. This has the advantages of more accurately evaluating recovered heat energy and the beneficial effect of improving the energy recovery efficiency of electric power equipment.

Description

Technical Field

[0001] The present invention relates to the field of electric energy data technology, and in particular to a method for energy recovery of electric power equipment based on multimodal processing. Background Art

[0002] With growing energy demand and increasing environmental protection requirements, energy recovery technology has been widely used in various power systems and equipment. Energy recovery systems can recover wasted energy (such as waste heat, vibration energy, and electromagnetic energy) from equipment, improving energy efficiency. Nonlinear high-frequency electromagnetic energy refers to energy fluctuations in high-frequency electromagnetic fields caused by the nonlinear characteristics of circuits or dielectrics (such as capacitance, inductance, and saturation of magnetic materials). This type of electromagnetic energy fluctuates dramatically, and under different frequency, power, and load conditions, the energy recovery efficiency and recovery strategy exhibit a strong nonlinear relationship. This nonlinear effect makes energy recovery systems difficult to accurately predict using simple linear models.

[0003] In conventional indoor electronic equipment, high-frequency electromagnetic energy is typically generated by the rapid switching of electronic devices such as motors, switching power supplies, and inverters. These devices primarily generate waste heat and often operate at frequencies higher than those of conventional power systems. Energy recovery systems typically rely on environmental data (such as temperature, humidity, and electromagnetic field strength) for predictions, and high-frequency electromagnetic signals can be mistakenly interpreted as valid signals. Unfiltered high-frequency noise can increase the strength of the "recovered" signal in the model, causing the system to overestimate the recovered energy. For example, high-frequency electromagnetic radiation can affect the system's temperature distribution or electromagnetic field environment, thereby affecting the efficiency of thermoelectric generators. If this radiation is not properly processed, the prediction model may erroneously estimate more waste heat than is actually achieved, resulting in an overestimation of the recovered heat energy. Most existing energy recovery systems use traditional linear models or empirical rules for energy recovery predictions, typically assuming a linear recovery process—that is, a direct and proportional relationship between the input signal and the recovered waste heat energy. Due to high-frequency electromagnetic noise and the nonlinear characteristics of circuits, these changes cannot be accurately modeled using traditional linear methods, which affects the energy efficiency of the system and makes it difficult for the system to effectively capture changes, resulting in a high accuracy in the prediction and evaluation of energy recovery when converting and recovering electrical energy. Summary of the Invention

[0004] The present invention provides an energy recovery method for electric power equipment based on multimodal processing, which solves the problem in the prior art that the recovery energy prediction result is too high due to the occurrence of high-frequency electromagnetic energy during the electric energy recovery process, resulting in poor evaluation accuracy.

[0005] The present invention is achieved through the following technical solutions:

[0006] A method for energy recovery of electric power equipment based on multimodal processing, the method comprising:

[0007] Step S1: monitoring and collecting load data, electromagnetic power values, and recovered heat energy values ​​generated by the target power equipment, and equally dividing the length of a single process execution time of the target power equipment into a number of working cycles;

[0008] Step S2: Based on the electromagnetic power value and the recovered heat energy value, detect whether the target power equipment has electromagnetic anomalies in each working cycle, screen out the working cycles with electromagnetic anomalies and mark them as abnormal cycles;

[0009] Step S3: Collecting load data of the target power equipment during the abnormal period, constructing a multimodal learning network based on the electromagnetic power value and the load data to generate constraint conditions, and using the constraint conditions to compensate and correct the recovered heat energy value during the abnormal period;

[0010] Step S4: collecting historical average values ​​of thermal energy for electric energy recovery of the target power equipment, setting a first reference thermal energy value in an increasing direction based on the historical average values ​​of thermal energy, and marking the recovered thermal energy value after compensation and correction as a correction thermal energy value;

[0011] Step S5: When the corrected thermal energy value is lower than the first reference thermal energy value, it is determined that the recovered thermal energy value in the current abnormal period has completed correction; when the corrected thermal energy value is higher than the first reference thermal energy value, step S3 is re-executed.

[0012] Furthermore, the electromagnetic anomaly detection process includes: setting a second reference thermal energy value along the increasing direction of the first reference thermal energy value, placing the recovered thermal energy value of the current working cycle, the first reference thermal energy value, and the second reference thermal energy value in the form of curve records in the same plane coordinate system, setting the horizontal axis to the time axis of the current process execution time, and setting an electromagnetic reference value for the electromagnetic power value;

[0013] When there is a time point on the time axis when the electromagnetic power value is lower than the electromagnetic reference value, it is determined that the heat energy recovery of the current target power equipment is not affected by electromagnetic radiation;

[0014] When there is a time point on the time axis at which the electromagnetic power value is higher than the electromagnetic reference value and the recovered heat energy value of the current working cycle is below the first reference heat energy value, it is determined that the heat energy recovery of the current target power equipment is not affected by electromagnetic radiation;

[0015] When there is a time point on the time axis where the electromagnetic power value is higher than the electromagnetic reference value, and the recovered heat energy value of the current working cycle is between the first reference heat energy value and the second reference heat energy value, it is determined that the heat energy recovery of the current target power equipment is affected by electromagnetic radiation, and the current working cycle is marked as an abnormal cycle;

[0016] When there is an electromagnetic power value higher than the electromagnetic reference value on the time axis, and the recovered heat energy value of the current working cycle is above the second reference heat energy value, it is determined that the heat energy recovery of the current target power equipment is affected by factors other than electromagnetic radiation, and the current correction process is terminated at this time.

[0017] Furthermore, an electromagnetic power average is calculated based on the electromagnetic power values ​​in the historical data of the target electric power equipment, and the electromagnetic power average is set as the electromagnetic reference value.

[0018] Furthermore, the standard deviation of the historical thermal energy mean is calculated, and the second reference thermal energy value is set to the historical thermal energy mean plus 2 times the standard deviation, and the first reference thermal energy value is set to the historical thermal energy mean plus 0.5 times the standard deviation.

[0019] Furthermore, the multimodal learning network is constructed based on a multilayer perceptron network; the process of generating the constraint conditions includes:

[0020] A multilayer perceptron network including an input layer, a hidden layer, and an output layer is constructed. The recovered heat energy value, electromagnetic power value, and load data of the target power equipment in the current abnormal period are normalized to generate a matrix vector. The input layer inputs the matrix vector of the electromagnetic power value and the load data as the input feature of the multilayer perceptron network to the hidden layer. The hidden layer uses the ReLU activation function to perform nonlinear calculation on the electromagnetic power value and the load data, and then transmits the calculation result to the output layer. An error correction factor representing the constraint condition is generated in the output layer, and the error correction factor is added to the matrix vector of the recovered heat energy value in a weighted form. The addition result is subjected to normalized inverse transformation processing, and the processing result of the normalized inverse transformation is marked as the corrected heat energy value.

[0021] Furthermore, the nonlinear calculation process of the hidden layer includes:

[0022] Let the function of input features be represented by X, let the number of hidden layers be a certain number, let i be the ordinal number of the hidden layer, and let the linear output and ReLU activation output of the hidden layer be represented by Z and A respectively, where i starts counting from layer 0; let the weight matrix of each hidden layer be represented by W i , let the bias term of each hidden layer be expressed as b i ; The input layer is represented as layer 0; set the activation output of layer 0 A0 = X,

[0023] Then the linear output Z of the i-th layer i The calculation formula is: ,

[0024] The ReLU activation output A of the i-th layer i The calculation formula is: ,

[0025] The ReLU activation output A calculated by the final layer of the hidden layer i Transfer to the output layer.

[0026] Furthermore, the calculation process of the output layer includes:

[0027] Let the total number of hidden layers be L, and the ReLU activation output transmitted to the output layer be A L ; Let the weight matrix of the output layer be represented as W L , let the bias term of the output layer be expressed as b L , and let the error correction factor be expressed as C,

[0028] Then the error correction factor C is calculated as: .

[0029] Furthermore, before normalizing the load data, Meyer wavelet transform is used to filter out electromagnetic mutation signals. The process includes:

[0030] The load data is decomposed into multiple scales, and different wavelet coefficients are obtained from the low-frequency band to the high-frequency band. The energy value of each wavelet coefficient is calculated, and the wavelet coefficient affected by the electromagnetic mutation is marked as the mutation signal coefficient.

[0031] The energy median of all wavelet coefficients is calculated, and the energy median is set as the wavelet threshold. The mutation signal coefficients are screened out using a soft threshold method based on the wavelet threshold, and the remaining wavelet coefficients are normalized.

[0032] Furthermore, the process of using the soft threshold method to screen out the mutation signal coefficient includes:

[0033] Use soft thresholding to process each mutation signal coefficient to generate a detail coefficient indicating that the soft thresholding process is completed; when the detail coefficient is greater than the wavelet threshold, the result after subtracting the wavelet threshold from the detail coefficient is used as the remaining retained wavelet coefficient for normalization; when the detail coefficient is less than the wavelet threshold, the coefficient is set to 0 and removed.

[0034] Furthermore, when performing multi-scale decomposition on the load data, the signal characteristics of the wavelet coefficients of each layer are extracted by calculating the statistical characteristics.

[0035] Compared with the existing technology, the present invention constructs a multimodal learning network based on electromagnetic power values ​​and load data by accurately detecting abnormal changes caused by high-frequency electromagnetic signals, so that the system can accurately capture the electromagnetic characteristics and energy recovery effects of the equipment under different load conditions. At the same time, it will automatically perform re-correction when the recovered thermal energy value is too high to ensure the prediction accuracy of the recovered energy. It has the advantages of more accurately evaluating the recovered thermal energy and the beneficial effect of improving the energy recovery efficiency of power equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0037] Figure 1 It is a flowchart of the present invention;

[0038] Figure 2 It is a schematic diagram of the process structure of the present invention. DETAILED DESCRIPTION

[0039] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention. Example

[0040] like Figure 1 As shown, this embodiment is a method for energy recovery of electric power equipment based on multimodal processing, the method comprising:

[0041] Step S1: monitoring and collecting load data, electromagnetic power values, and recovered heat energy values ​​generated by the target power equipment, and equally dividing the length of a single process execution time of the target power equipment into a number of working cycles;

[0042] Step S2: Based on the electromagnetic power value and the recovered heat energy value, detect whether the target power equipment has electromagnetic anomalies in each working cycle, screen out the working cycles with electromagnetic anomalies and mark them as abnormal cycles;

[0043] Step S3: Collecting load data of the target power equipment during the abnormal period, constructing a multimodal learning network based on the electromagnetic power value and the load data to generate constraint conditions, and using the constraint conditions to compensate and correct the recovered heat energy value during the abnormal period;

[0044] Step S4: collecting historical average values ​​of thermal energy for electric energy recovery of the target power equipment, setting a first reference thermal energy value in an increasing direction based on the historical average values ​​of thermal energy, and marking the recovered thermal energy value after compensation and correction as a correction thermal energy value;

[0045] Step S5: When the corrected thermal energy value is lower than the first reference thermal energy value, it is determined that the recovered thermal energy value in the current abnormal period has completed correction; when the corrected thermal energy value is higher than the first reference thermal energy value, step S3 is re-executed.

[0046] Furthermore, as a feasible implementation method, the electromagnetic anomaly detection process includes: setting a second reference thermal energy value along the increasing direction of the first reference thermal energy value, placing the recovered thermal energy value of the current working cycle, the first reference thermal energy value, and the second reference thermal energy value in the form of curve records in the same plane coordinate system, setting the horizontal axis to the time axis of the current process execution time, and setting an electromagnetic reference value for the electromagnetic power value;

[0047] When there is a time point on the time axis when the electromagnetic power value is lower than the electromagnetic reference value, it is determined that the heat energy recovery of the current target power equipment is not affected by electromagnetic radiation;

[0048] When there is a time point on the time axis at which the electromagnetic power value is higher than the electromagnetic reference value and the recovered heat energy value of the current working cycle is below the first reference heat energy value, it is determined that the heat energy recovery of the current target power equipment is not affected by electromagnetic radiation;

[0049] When there is a time point on the time axis where the electromagnetic power value is higher than the electromagnetic reference value, and the recovered heat energy value of the current working cycle is between the first reference heat energy value and the second reference heat energy value, it is determined that the heat energy recovery of the current target power equipment is affected by electromagnetic radiation, and the current working cycle is marked as an abnormal cycle;

[0050] When there is an electromagnetic power value higher than the electromagnetic reference value on the time axis, and the recovered heat energy value of the current working cycle is above the second reference heat energy value, it is determined that the heat energy recovery of the current target power equipment is affected by factors other than electromagnetic radiation, and the current correction process is terminated at this time.

[0051] The target power equipment is one that is performing heat recovery and is subject to high-frequency electromagnetic signal interference. This electromagnetic signal interference can cause significant deviations in the heat recovery data for this power equipment, necessitating data correction. Load data refers to various types of information related to the load borne by the power equipment during operation. This data can reflect the equipment's energy consumption, operating status, and load variations under different operating conditions. In specific applications, this load data may include operating parameters such as the target power equipment's power, current and voltage, operating temperature, power factor, and operating frequency. The electromagnetic power value refers to the power corresponding to the electromagnetic field strength and associated electromagnetic fluctuations generated by the target power equipment during operation. It represents the energy transfer rate generated by the interaction of the electric and magnetic fields within the electromagnetic field. In specific applications, the electromagnetic power value of the target power equipment can be directly measured using an electromagnetic field sensor to directly measure the electromagnetic field strength generated by the equipment, thereby calculating the power of the electromagnetic waves by measuring the strength of the electric and magnetic fields. Alternatively, electromagnetic power meters can be installed at the input and output terminals of the power equipment to directly measure the power of the electromagnetic waves. The recovered thermal energy value refers to the amount of waste heat or other forms of energy generated by the equipment during operation that is effectively captured and converted into usable energy by the recovery system. In practical applications, a heat flow meter can be used to directly measure the heat transfer through an object's surface or material per unit time to obtain the recovered heat value, or a temperature sensor such as a thermocouple or RTD sensor can be used to monitor and obtain the recovered heat value. The single process execution time of the target power equipment refers to the time required for the target power equipment to complete a specific task or process. Electromagnetic anomaly detection means that if significant abnormal fluctuations in the electromagnetic power value and recovered heat value occur within a certain working cycle, the system will determine that an electromagnetic anomaly exists in that working cycle. These abnormal fluctuations are likely caused by nonlinear electromagnetic effects within the equipment. Constraints are used to guide the correction of recovered heat values ​​within the multimodal learning network. When the system predicts a recovered heat value that is too high, the generated constraints will set an upper limit for the recovered heat value to avoid overestimation. When the recovered heat value is too low, the constraints can guide the model to adjust the predicted value to better reflect actual conditions. The redundant configuration of the first reference heat value provides a safety margin to avoid deviations caused by model errors or abnormal data, ensuring that the heat value remains within a reasonable range.

[0052] Furthermore, during electromagnetic anomaly detection, when there is a time point on the time axis where the electromagnetic power value is below the electromagnetic reference value, this condition indicates that the electromagnetic power intensity at the current time point is below the preset electromagnetic reference value, meaning that the electromagnetic radiation at the device at that moment is weak or has almost no radiation impact, indicating that the device's heat recovery is not significantly affected by electromagnetic radiation. When there is a time point on the time axis where the electromagnetic power value is above the electromagnetic reference value, and the recovered heat energy value of the current operating cycle is below the first reference heat energy value, it is determined that the electromagnetic radiation in the current target power device, although reaching an abnormal indicator level, has no significant impact on the target power device's heat recovery process. When there is a time point on the time axis where the electromagnetic power value is above the electromagnetic reference value, and the recovered heat energy value of the current operating cycle is between the first reference heat energy value and the second reference heat energy value, it is determined that the recovered heat energy value of the target power device at that time was affected by electromagnetic radiation. Because it does not exceed the second reference heat energy value, electromagnetic radiation is determined to be the primary factor affecting the elevated heat energy data. When there is an electromagnetic power value on the time axis that is higher than the electromagnetic reference value, and the recovered heat energy value of the current working cycle is above the second reference heat energy value, since the recovered heat energy value at this time exceeds the range for judging that electromagnetic radiation is the main influence, it is determined that the target power equipment is affected by energy factors other than high-frequency electromagnetic radiation, indicating that the current detection process lacks accuracy, the current correction process is terminated and other data correction methods are adopted.

[0053] Among them, the electromagnetic power mean is calculated based on the electromagnetic power value in the historical data of the target power equipment, and the electromagnetic power mean is set as the electromagnetic reference value; the standard deviation of the historical thermal energy mean is calculated, and the second reference thermal energy value is set as the historical thermal energy mean plus 2 times the standard deviation, and the first reference thermal energy value is set as the historical thermal energy mean plus 0.5 times the standard deviation.

[0054] Setting the mean electromagnetic power as the electromagnetic reference value accurately describes the electromagnetic power fluctuation range of the device under normal operating conditions. Compared to single measurements or assumed values, the mean value more stably reflects the typical operating characteristics of the device. Setting the first reference thermal energy value to the historical mean thermal energy value plus 0.5 standard deviation is used to capture small fluctuations. This serves as the lower threshold for normal operation and is adaptable to minor fluctuations in the device. This ensures the system maintains high sensitivity to small fluctuations, ensuring that fluctuations within the normal range are not mistakenly identified as abnormalities, and can detect even slight deviations from normal expectations. Setting the second reference thermal energy value to the historical mean thermal energy value plus 2 standard deviations is used to identify significantly abnormal fluctuations. This means that, under a normal distribution, approximately 95% of the data will fall within the range of ±2 standard deviations of the mean. If a data point exceeds this range of 2 standard deviations, it is considered unlikely to occur and is a significant anomaly, allowing the second reference thermal energy value to tolerate larger fluctuations. If the recovered thermal energy value exceeds this value, it can be determined that the fluctuation is likely caused by external factors such as electromagnetic radiation and thus marked as an abnormal period. This "double standard deviation multiple" setting method helps the system maintain efficient and accurate anomaly detection capabilities when facing fluctuations of different amplitudes that may occur during the equipment's heat recovery process.

[0055] Furthermore, as a feasible implementation method, the multimodal learning network is constructed based on a multilayer perceptron network; the process of generating the constraint conditions includes:

[0056] A multilayer perceptron network including an input layer, a hidden layer, and an output layer is constructed. The recovered heat energy value, electromagnetic power value, and load data of the target power equipment in the current abnormal period are normalized to generate a matrix vector. The input layer inputs the matrix vector of the electromagnetic power value and the load data as the input feature of the multilayer perceptron network to the hidden layer. The hidden layer uses the ReLU activation function to perform nonlinear calculation on the electromagnetic power value and the load data, and then transmits the calculation result to the output layer. An error correction factor representing the constraint condition is generated in the output layer, and the error correction factor is added to the matrix vector of the recovered heat energy value in a weighted form. The addition result is subjected to normalized inverse transformation processing, and the processing result of the normalized inverse transformation is marked as the corrected heat energy value.

[0057] The multilayer perceptron (MLP) network, a classic neural network architecture, can effectively process a variety of input features and capture complex feature relationships through multiple nonlinear computational layers. In this embodiment, the computational layers are the hidden layers of the MLP network. The output layer normalizes the electromagnetic power and load data of the target power equipment and passes them as input features to the hidden layer. The hidden layer uses the Rectified Linear Unit (RLU) activation function to perform nonlinear mapping on the electromagnetic power and load data. The ReLU function effectively avoids the vanishing gradient problem and offers advantages such as high computational efficiency and fast convergence, enabling the network to better learn complex nonlinear relationships. The output layer generates an error correction factor, a key parameter used to correct errors in the recovered heat energy value prediction. The implementation process is as follows: the matrix vectors of electromagnetic power and load data in the input layer are passed as input features to the hidden layer. The hidden layer performs nonlinear computations on these data using the ReLU activation function, capturing the complex interactions between the electromagnetic power and load data. The computational results are then passed to the output layer to generate the error correction factor. The error correction factor generated by the output layer is used to correct the recovered heat energy value. The correction process first adds the recovered heat value and the error correction factor, then performs a normalized inverse transformation to obtain the corrected recovered heat value, known as the corrected heat value. By training the network to generate the error correction factor, the system automatically adjusts the recovered heat value prediction based on the electromagnetic power and load data of the target power equipment during abnormal periods, reducing the need for manual intervention and improving intelligence.

[0058] Furthermore, as a feasible implementation method, the nonlinear calculation process of the hidden layer includes:

[0059] Let the function of input features be represented by X, let the number of hidden layers be a certain number, let i be the ordinal number of the hidden layer, and let the linear output and ReLU activation output of the hidden layer be represented by Z and A respectively, where i starts counting from layer 0; let the weight matrix of each hidden layer be represented by W i , let the bias term of each hidden layer be expressed as b i ; The input layer is represented as layer 0; set the activation output of layer 0 A0 = X,

[0060] Then the linear output Z of the i-th layer i The calculation formula is: ,

[0061] The ReLU activation output A of the i-th layer i The calculation formula is: ,

[0062] The ReLU activation output A calculated by the final layer of the hidden layeri Transfer to the output layer.

[0063] The input features are represented by vectors as X. This process ensures that the multimodal data of electromagnetic power values ​​and load data can be effectively input into the network, which becomes the basis for model training. Then the linear output Z of the i-th layer is i Represented by the weight matrix W i and the bias term b i The activation output A of the previous layer i-1 Perform linear transformation and use the weighted sum of the input data plus the bias to obtain the output of the intermediate layer. The ReLU activation output A i Indicates linear output Z i The result is transformed nonlinearly using the ReLU activation function. The ReLU function is a commonly used activation function that maps negative values ​​to zero and keeps positive values ​​unchanged. It helps introduce nonlinear factors and enables the neural network to handle more complex patterns and relationships. The ReLU activation output of each layer is A i Will be passed to the next layer until the output A of the last layer i Passed to the output layer. Through the multi-layer perceptron network, the input features are transformed layer by layer nonlinearly, and the model can better capture the complex relationships in the data. Among them, the ReLU activation function can effectively avoid the gradient vanishing problem. Compared with the Sigmoid or Tanh function, the ReLU activation function can maintain positive values ​​and is simple and efficient. It will not cause the training process to be slow or fail due to too small gradients during the back propagation process. The weight matrix W i This enables the network to extract key features from the data and optimize the output prediction based on these features. i It provides additional flexibility to each layer, ensuring that the network can handle input data that is not centered at the origin.

[0064] Furthermore, as a feasible implementation method, the calculation process of the output layer includes:

[0065] Let the total number of hidden layers be L, and the ReLU activation output transmitted to the output layer be A L ; Let the weight matrix of the output layer be represented by W L , let the bias term of the output layer be expressed as b L , and let the error correction factor be expressed as C,

[0066] Then the error correction factor C is calculated as: .

[0067] In a multi-layer perceptron network, after the nonlinear calculation of the L hidden layers, the activation output of the last layer is A L, which is the final feature representation generated by the nonlinear transformation of the layer-by-layer ReLU activation function. The connection between the output layer and the previous layer is through the weight matrix W of the output layer. L and the bias term b L Adjust the weight matrix W of the output layer L Determines each input feature (i.e., A L The influence of each element in on the output, and the bias term b L Used to adjust the final result. The output layer is passed through the weight matrix W L and the bias term b L Calculate a correction factor C. This factor is used to adjust the prediction error of the recovered heat value so that the prediction result is more consistent with the actual situation. During the training process, the calculation of the error correction factor C helps to continuously optimize the weight matrix W through the back propagation algorithm. L and the bias term b L , thereby improving the model's ability to predict the operating status of complex power equipment. By optimizing these parameters, the network can gradually approach the actual situation of recovered thermal energy value.

[0068] Furthermore, before normalizing the load data, Meyer wavelet transform is used to filter out electromagnetic mutation signals. The process includes:

[0069] The load data is decomposed into multiple scales, and different wavelet coefficients are obtained from the low-frequency band to the high-frequency band. The energy value of each wavelet coefficient is calculated, and the wavelet coefficient affected by the electromagnetic mutation is marked as the mutation signal coefficient.

[0070] The energy median of all wavelet coefficients is calculated, and the energy median is set as the wavelet threshold. The mutation signal coefficients are screened out using a soft threshold method based on the wavelet threshold, and the remaining wavelet coefficients are normalized.

[0071] Each wavelet coefficient represents the characteristic intensity of the signal at a specific scale, and the energy value can measure the intensity of signal activity at that scale. By calculating the energy value of each wavelet coefficient, parts with large energy changes can be identified, which are usually related to electromagnetic mutations. In order to avoid the influence of extreme values, this embodiment uses the median of the energy value as the wavelet threshold. The median is a statistic that is insensitive to outliers and can effectively avoid erroneous judgments caused by certain extremely large or extremely small energy values. The soft threshold method is a commonly used signal denoising method that can suppress mutation signals by setting a threshold. When the energy of the wavelet coefficient exceeds the threshold, the correlation coefficient will be adjusted according to the threshold so that the mutation signal is effectively suppressed. Compared with the hard threshold, the soft threshold processing can retain more signal information and make the processed signal smoother, avoiding the distortion that may occur after the mutation signal processing.

[0072] Furthermore, the process of using the soft threshold method to screen out the mutation signal coefficient includes:

[0073] Each mutation signal coefficient is processed using a soft threshold to generate a detail coefficient indicating that the soft threshold processing is completed; when the detail coefficient is greater than the wavelet threshold, the result after subtracting the wavelet threshold from the detail coefficient is used as the remaining retained wavelet coefficient for normalization processing; when the detail coefficient is less than the wavelet threshold, the coefficient is set to 0 and removed; when the load data is subjected to multi-scale decomposition, the signal characteristics of the wavelet coefficients of each layer are extracted by calculating statistical features.

[0074] The load data is subjected to a wavelet transform to generate wavelet coefficients. The wavelet coefficients at each layer represent data characteristics at different scales. The high-frequency components of the wavelet coefficients (i.e., the coefficients of sudden changes in the signal) after soft thresholding are represented as detail coefficients. These high-frequency components typically represent signal details, such as noise or electromagnetic fluctuations. When the detail coefficient is greater than the wavelet threshold, the result of subtracting the threshold from the detail coefficient is used as the remaining retained wavelet coefficient. This smoothes the signal details and preserves strong signal features. When the detail coefficient is less than the wavelet threshold, it is set to 0, indicating that these coefficients represent noise or sudden changes in the signal and are completely removed. Soft thresholding of the detail coefficients retains coefficients with strong signal strength, i.e., those whose absolute values ​​exceed the wavelet threshold. These retained coefficients reflect the true fluctuation information in the signal and effectively preserve important signal features. When the detail coefficient is less than the wavelet threshold, the system sets it to 0, removing high-frequency components caused by electromagnetic fluctuations or noise. This process effectively suppresses high-frequency noise caused by electromagnetic radiation, device switching, and other sources. This embodiment can effectively suppress high-frequency noise caused by electromagnetic mutations, ensuring that the prediction model of recovered thermal energy value will not be interfered with by these abnormal signals.

[0075] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for energy recovery of electric power equipment based on multimodal processing, characterized in that: The method includes: Step S1: monitoring and collecting load data, electromagnetic power values, and recovered heat energy values ​​generated by the target power equipment, and equally dividing the length of a single process execution time of the target power equipment into a number of working cycles; Step S2: Based on the electromagnetic power value and the recovered heat energy value, detect whether the target power equipment has electromagnetic anomalies in each working cycle, screen out the working cycles with electromagnetic anomalies and mark them as abnormal cycles; Step S3: Collecting load data of the target power equipment during the abnormal period, constructing a multimodal learning network based on the electromagnetic power value and the load data to generate constraint conditions, and using the constraint conditions to compensate and correct the recovered heat energy value during the abnormal period; Step S4: collecting historical average values ​​of thermal energy for electric energy recovery of the target power equipment, setting a first reference thermal energy value in an increasing direction based on the historical average values ​​of thermal energy, and marking the recovered thermal energy value after compensation and correction as a correction thermal energy value; Step S5: When the corrected heat energy value is lower than the first reference heat energy value, it is determined that the recovered heat energy value in the current abnormal period has been corrected. When the corrected heat energy value is higher than the first reference heat energy value, step S3 is executed again. The multimodal learning network is constructed based on a multilayer perceptron network; the process of generating the constraint conditions includes: Constructing a multilayer perceptron network comprising an input layer, a hidden layer, and an output layer, normalizing the recovered heat energy value, electromagnetic power value, and load data of the target power equipment in the current abnormal period to generate a matrix vector, the input layer inputting the matrix vector of the electromagnetic power value and the load data as input features of the multilayer perceptron network to the hidden layer, performing a nonlinear calculation on the electromagnetic power value and the load data using a ReLU activation function in the hidden layer, and transmitting the calculation result to the output layer, generating an error correction factor representing the constraint condition in the output layer, adding the error correction factor to the matrix vector of the recovered heat energy value in a weighted form, performing a normalized inverse transformation on the addition result, and marking the normalized inverse transformation result as the corrected heat energy value; Before normalizing the load data, Meyer wavelet transform is used to filter out electromagnetic mutation signals. The process includes: The load data is decomposed into multiple scales, and different wavelet coefficients are obtained from the low-frequency band to the high-frequency band. The energy value of each wavelet coefficient is calculated, and the wavelet coefficient affected by the electromagnetic mutation is marked as the mutation signal coefficient. The energy median of all wavelet coefficients is calculated, and the energy median is set as the wavelet threshold. The mutation signal coefficients are screened out using a soft threshold method based on the wavelet threshold, and the remaining wavelet coefficients are normalized.

2. The method for energy recovery of electric power equipment based on multimodal processing according to claim 1, characterized in that: The electromagnetic anomaly detection process includes: setting a second reference thermal energy value along the direction of increasing the value of the first reference thermal energy value, placing the recovered thermal energy value of the current working cycle, the first reference thermal energy value, and the second reference thermal energy value in the form of a curve record in the same plane coordinate system, setting the horizontal axis to the time axis of the current process execution time, and setting an electromagnetic reference value for the electromagnetic power value; When there is a time point on the time axis when the electromagnetic power value is lower than the electromagnetic reference value, it is determined that the heat energy recovery of the current target power equipment is not affected by electromagnetic radiation; When there is a time point on the time axis at which the electromagnetic power value is higher than the electromagnetic reference value and the recovered heat energy value of the current working cycle is below the first reference heat energy value, it is determined that the heat energy recovery of the current target power equipment is not affected by electromagnetic radiation; When there is a time point on the time axis where the electromagnetic power value is higher than the electromagnetic reference value, and the recovered heat energy value of the current working cycle is between the first reference heat energy value and the second reference heat energy value, it is determined that the heat energy recovery of the current target power equipment is affected by electromagnetic radiation, and the current working cycle is marked as an abnormal cycle; When there is an electromagnetic power value higher than the electromagnetic reference value on the time axis, and the recovered heat energy value of the current working cycle is above the second reference heat energy value, it is determined that the heat energy recovery of the current target power equipment is affected by factors other than electromagnetic radiation, and the current correction process is terminated at this time.

3. The method for energy recovery of electric power equipment based on multimodal processing according to claim 2, characterized in that: An electromagnetic power average is calculated based on electromagnetic power values ​​in historical data of the target electric power equipment, and the electromagnetic power average is set as an electromagnetic reference value.

4. The method for energy recovery of electric power equipment based on multimodal processing according to claim 2, characterized in that: Calculate the standard deviation of the historical thermal energy mean, set the second reference thermal energy value to the historical thermal energy mean plus 2 times the standard deviation, and set the first reference thermal energy value to the historical thermal energy mean plus 0.5 times the standard deviation.

5. The method for energy recovery of electric power equipment based on multimodal processing according to claim 1, characterized in that: The nonlinear calculation process of the hidden layer includes: Let the function of input features be represented by X, let the number of hidden layers be a certain number, let i be the ordinal number of the hidden layer, and let the linear output and ReLU activation output of the hidden layer be represented by Z and A respectively, where i starts counting from layer 0; let the weight matrix of each hidden layer be represented by W i , let the bias term of each hidden layer be expressed as b i ; The input layer is represented as layer 0; set the activation output of layer 0 A0 = X, Then the linear output Z of the i-th layer i The calculation formula is: , The ReLU activation output A of the i-th layer i The calculation formula is: , The ReLU activation output A calculated by the final layer of the hidden layer i Transfer to the output layer.

6. The method for energy recovery of electric power equipment based on multimodal processing according to claim 5, characterized in that: The calculation process of the output layer includes: Let the total number of hidden layers be L, and the ReLU activation output transmitted to the output layer be A L ; Let the weight matrix of the output layer be represented by W L , let the bias term of the output layer be expressed as b L , and let the error correction factor be expressed as C, Then the error correction factor C is calculated as: .

7. The method for energy recovery of electric power equipment based on multimodal processing according to claim 1, characterized in that: The process of using the soft threshold method to screen out the mutation signal coefficient includes: Use soft thresholding to process each mutation signal coefficient to generate a detail coefficient indicating that the soft thresholding process is completed; when the detail coefficient is greater than the wavelet threshold, the result after subtracting the wavelet threshold from the detail coefficient is used as the remaining retained wavelet coefficient for normalization; when the detail coefficient is less than the wavelet threshold, the coefficient is set to 0 and removed.

8. The method for energy recovery of electric power equipment based on multimodal processing according to claim 1, characterized in that: When performing multi-scale decomposition on the load data, the signal characteristics of the wavelet coefficients of each layer are extracted by calculating the statistical characteristics.

Images