Semi-supervised training method and device for energy heterogeneous data credible compensation AI model

Abstract

The invention relates to a semi-supervised training method and device for an energy heterogeneous data credible compensation AI model. The method comprises the steps that unified time sequence formatting, rule verification based on equipment rated parameters and transient characteristics, Bayesian weight distribution and time sequence compensation based on medium characteristics and equipment dynamic coupling lag response characteristics are sequentially carried out on multi-source heterogeneous data; pre-training a teacher model under small sample labeling; performing online distillation on the student model under the unlabeled sample; and teacher and student double-model fusion reasoning is carried out. The device comprises a unified formatting module, a physical constraint cleaning module, a data quality weighting module, a time sequence lag compensation module, a teacher model generation module, an online distillation training module and a reasoning prediction module. According to the method, the data value is maximized through semi-supervised training under low annotation, lightweight design is adapted to edge deployment, the credibility is guaranteed through physical constraints, and more efficient and real-time energy AI is realized.

Description

Technical Field

[0001] This invention belongs to the field of industrial control energy and relates to a semi-supervised training method, device and computer equipment for a trusted AI model for energy heterogeneous data compensation. Background Technology

[0002] Semi-supervised training refers to the joint modeling using a small number of labeled samples and a large number of unlabeled running data. Through mechanisms such as pseudo-label generation and consistency constraints, it reduces labeling costs and improves the model's generalization ability.

[0003] Because energy data sources cover a variety of sensing and information systems such as SCADA systems and smart meters, and the data types include time-series loads, voltage and current waveforms, with sampling frequencies ranging from milliseconds to hours, energy data often exhibits fragmentation in semantic, temporal, and quality dimensions, presenting multi-source heterogeneous characteristics.

[0004] The application of energy AI models has become a trend and an important technological path to promote the digital transformation of the energy industry. If models can be built by deeply integrating the operating characteristics of energy equipment with the features of multi-source data, it can provide key support for scenarios such as high-proportion renewable energy grid connection, real-time load forecasting, and intelligent operation and maintenance of equipment.

[0005] Currently, existing technology discloses a method and system for constructing a power time-series forecast dataset (application number CN202511440207.5). This method collects power load, weather, and multi-source common time-series data, aligns them with timestamps, detects statistical anomalies, and uses dynamic weighted sampling to construct a standardized heterogeneous dataset for model training. This approach uses similarity indicators to filter strongly correlated data, enhancing the representativeness of the training set; the dynamic sampling weighting strategy balances the contributions of data from different sources, avoiding sampling bias. However, this method is highly sensitive to data quality and relies on statistical detection and manually set cleaning rules. For lag effects in time-series data, such as the delayed impact of meteorological factors on load, a fixed sliding window or historical mean compensation is used. In cases of poor data quality or seasonal transitions, frequent adjustments to rule thresholds are required, increasing maintenance costs and reliance on expert experience.

[0006] Therefore, how to build a data quality self-repair and hysteresis adaptive compensation mechanism in the training of energy AI models, which can dynamically adapt to the temporal characteristics and operating condition drift of multi-source heterogeneous data, and get rid of the strong dependence on manual annotation and expert parameter tuning, and achieve the technical goal of "low governance cost, few available samples, and high reliability inference", has become a core problem that urgently needs to be solved. Summary of the Invention

[0007] In order to solve the technical problems in the background art, the present invention provides a semi-supervised training method, apparatus and computer equipment for a reliable compensation AI model for heterogeneous energy data.

[0008] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: The first aspect provides a semi-supervised training method for a reliable compensation AI model for heterogeneous energy data, the method comprising the following steps: The scattered raw data is unified in terms of structure, semantics, and temporal sequence to obtain formatted data; Physical rule verification is performed on outliers in formatted data based on the equipment's rated parameters and transient characteristics. Dynamically assign Bayesian weights to the validated data to obtain weighted data with confidence labels; Based on the hysteresis response characteristics of the coupling between the medium properties and the equipment dynamics, the weighted data is time-series misaligned to obtain time-compensated weighted data. A small number of samples were extracted from the time-compensated weighted data to train the DistilBERT teacher model; The DistilBERT teacher model generates pseudo-labels from formatted data to drive online distillation training of the LSTM student model, and updates the teacher parameters based on the accumulated actual running error. During the inference phase, the trained DistilBERT teacher model extracts static semantic features, and the trained LSTM student model extracts dynamic temporal features. The prediction results are then output after being weighted and fused by unidimensional attention.

[0009] Secondly, a semi-supervised training device for a reliable compensation AI model for heterogeneous energy data is provided, the device comprising: The unified formatting module is used to unify the structure, semantics, and temporal sequence of scattered raw data to obtain formatted data; The physical constraint cleaning module is used to perform physical rule verification on outliers in formatted data based on the equipment's rated parameters and transient characteristics. The data quality weighting module is used to dynamically assign Bayesian weights to the verified data to obtain weighted data with confidence labels. The timing lag compensation module is used to perform timing misalignment alignment on the weighted data based on the lag response characteristics of the medium characteristics and the dynamic coupling of the equipment, so as to obtain the timing-compensated weighted data. The teacher model generation module is used to extract a small number of samples from the time-compensated weighted data to train the DistilBERT teacher model. The online distillation training module is used to generate pseudo-labels from formatted data to drive the online distillation training of the LSTM student model from the DistilBERT teacher model, and to update the teacher parameters based on the actual running error. The inference prediction module is used in the inference stage. The trained DistilBERT teacher model extracts static semantic features, and the trained LSTM student model extracts dynamic temporal features. The prediction results are then output after being weighted and fused by unidimensional attention.

[0010] Thirdly, a computer device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the steps of a semi-supervised training method for a trusted AI model for energy heterogeneous data compensation as described in any of the preceding claims.

[0011] The beneficial effects of this invention are: (1) This invention replaces fixed statistical rules with dynamic cleaning driven by physical constraints. Combined with Bayesian weighted formatting, the system can autonomously adapt to the quality fluctuations of multi-source heterogeneous data without frequent manual parameter tuning. The time lag compensation is based on the dynamic calculation of lag time based on the dynamic coupling characteristics of equipment, which gets rid of the rigid limitation of the traditional sliding window and has stronger universality. (2) This invention adopts semi-supervised pre-training and online distillation training, which requires only a small number of labeled samples. Combined with the feedback of actual running error, it continuously utilizes unlabeled data, effectively alleviating the problem of label scarcity and forming a closed loop of data self-repair and model self-optimization. (3) When reasoning, the DistilBERT teacher model and the lightweight LSTM student model are integrated into an architecture. The single-dimensional attention dynamic weighting ensures both accuracy and meets the computing power constraints of the edge. It constructs a complete closed-loop system from data governance, online learning to lightweight reasoning, which solves the core contradictions of existing technologies such as fragmentation, high governance costs, and difficulty in balancing credibility and real-time performance. It is more in line with the needs of large energy AI models in scenarios with few samples and strong time delays. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the semi-supervised training method for the AI ​​model for reliable compensation of heterogeneous energy data provided in Embodiment 1 of the present invention.

[0014] Figure 2 This is a structural diagram of the timing misalignment mapping table generated in Embodiment 1 of the present invention.

[0015] Figure 3This is a schematic diagram of the timing misalignment alignment principle driven by the hysteresis response characteristics of the coupling between medium properties and equipment dynamics in Embodiment 1 of the present invention.

[0016] Figure 4 This is a schematic diagram of the semi-supervised training device for the AI ​​model of energy heterogeneous data credibility compensation provided in Embodiment 2 of the present invention.

[0017] Figure 5 This is a schematic diagram of the structure of the electronic device provided in Embodiment 3 of the present invention.

[0018] The attached diagram lists the components represented by each number as follows: 2001, Unified Formatting Module; 2002, Physical Constraint Cleaning Module; 2003, Data Quality Weighting Module; 2004, Time Lag Compensation Module; 2005, Teacher Model Generation Module; 2006, Online Distillation Training Module; 2007, Inference Prediction Module; 310, Processor; 320, Communication Interface; 330, Memory; 340, Communication Bus. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] Example 1

[0021] Currently, existing technology discloses a method and system for constructing a power time-series forecast dataset (application number CN202511440207.5). This method collects power load, weather, and multi-source common time-series data, aligns them with timestamps, detects statistical anomalies, and uses dynamic weighted sampling to construct a standardized heterogeneous dataset for model training. This approach uses similarity indicators to filter strongly correlated data, enhancing the representativeness of the training set; the dynamic sampling weighting strategy balances the contributions of data from different sources, avoiding sampling bias. However, this method is highly sensitive to data quality and relies on statistical detection and manually set cleaning rules. For lag effects in time-series data, such as the delayed impact of meteorological factors on load, a fixed sliding window or historical mean compensation is used. In cases of poor data quality or seasonal transitions, frequent adjustments to rule thresholds are required, increasing maintenance costs and reliance on expert experience.

[0022] Therefore, how to build a data quality self-repair and hysteresis adaptive compensation mechanism in the training of energy AI models, which can dynamically adapt to the temporal characteristics and operating condition drift of multi-source heterogeneous data, and get rid of the strong dependence on manual annotation and expert parameter tuning, and achieve the technical goal of "low governance cost, few available samples, and high reliability inference", has become a core problem that urgently needs to be solved.

[0023] To address the aforementioned issues, this invention provides a semi-supervised training method for a reliable compensation AI model for heterogeneous energy data. Figure 1 This is a schematic diagram of the semi-supervised training method for the AI ​​model for reliable compensation of heterogeneous energy data provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this method includes: Step S101: The scattered raw data is unified in terms of structure, semantics and temporal sequence to obtain formatted data.

[0024] Preferably, step S101 further includes: converting the scattered raw data into formatted data with unified structure, time sequence, and semantics, wherein the structure encoding rule is UTF-8; the time sequence unified timestamp is year-month-day-hour-minute-second; and the semantics are unified as energy consumption / kilowatt-hour.

[0025] Understandably, energy data comes from a wide range of sources, and this example captures raw data generated by various devices and systems, including power SCADA systems, photovoltaic inverters, and equipment PLCs.

[0026] It is also understandable that this embodiment uses the UTF-8 encoding structure, which supports multiple language character sets, ensuring seamless compatibility of data from different regions; it is also backward compatible with ASCII, facilitating the smooth integration of data generated by earlier systems into modern systems. UTF-8 employs an efficient variable-length encoding method, suitable for processing text data containing a large number of non-ASCII characters, and is widely used in modern computing systems, ensuring efficient data transmission and interoperability between different systems. Compared to other encoding methods, UTF-8 performs better in terms of universality, compatibility, and standardization, providing a reliable encoding foundation for the integration and analysis of energy data.

[0027] This embodiment adopts a unified year-month-day-hour-minute-second timestamp format. In energy data analysis, especially in complex time series analyses such as load forecasting and equipment fault diagnosis, time consistency accurate to the second is crucial. A unified time series format not only simplifies the data processing workflow and reduces the workload of complex time format conversions before analysis, but also provides a reliable time reference for various time series analyses, making data analysis more accurate and efficient.

[0028] This embodiment uses the unified semantic unit of energy consumption / kWh. By standardizing energy consumption data from different energy sources, such as electricity, steam, and gas, under the same measurement standard, this allows for direct comparison and analysis of these data, enhancing their comparability and analytical value. This unification not only simplifies the data integration process and avoids complex unit conversions but also provides a consistent data foundation for cross-domain applications. Whether for energy management, energy conservation optimization, or carbon emission calculation, energy consumption / kWh, as a universal energy measurement unit, provides unified and reliable data support, promoting intelligent and refined management of energy systems.

[0029] Step S102: Perform physical rule verification on the abnormal values ​​of the formatted data based on the rated parameters and transient characteristics of the equipment.

[0030] Preferably, step S102 further includes: verifying abnormal values ​​of formatted data according to the rated parameters of the equipment; if the values ​​exceed the limits, the transient physical model of the equipment is called to distinguish between instrument jumps and actual overloads; jump data is corrected by reverse calculation from the transient physical model of the equipment; and overload data is passed after being marked with an overload flag; if the values ​​do not exceed the limits, the data is passed directly.

[0031] In this embodiment, the physical constraint-driven cleaning rules embed the transient characteristics of multiple energy media, such as electrical and thermodynamic, into the decision flow. When a sensor exceeds its rated parameter threshold, it is not directly rejected. Instead, the system calls the corresponding medium's equipment transient physical model to determine whether it is an instrument jump or a sudden change in actual operating conditions. For example, for steam flow, sensor drift is verified based on the pipeline pressure drop-flow characteristic (Bernoulli equation). In this way, the system can distinguish between "data errors" and "physical truths," maintaining a low false positive rate even in scenarios with few samples. Furthermore, it adaptively identifies equipment aging through online parameter identification, avoiding frequent manual recalibration.

[0032] It is important to note that for multi-source data such as electricity, steam, and gas, the selection of the equipment transient physical model should follow the principle of media characteristics-equipment coupling: Power system: Select RLC transient model or motor starting current curve, and distinguish between meter jump (instantaneous spike without physical inertia) and real overload (which conforms to the e﹣t / τ decay law) by analyzing the rising slope and decay time constant of the current waveform. Steam / gas system: A hydraulic-thermal coupling model of the pipeline network or the transient flow curve of the burner is used. The nonlinear relationship between pressure and flow, such as the pressure wave propagation speed of the steam pipeline network, is utilized to determine sensor drift (sudden pressure change but no flow response) or actual leakage (abnormal pressure-flow synchronization). Therefore, the selection of the equipment transient physical model only needs to meet the above principles, and this embodiment does not impose specific limitations.

[0033] It is worth noting that the jump data corrected by the device's transient physical model are retained in the training set, while the actual overload data are labeled as "high-confidence extreme samples." Both are then input into subsequent Bayesian weighted formatting and online distillation. This ensures that even with scarce labeled samples, the system can still generate reliable pseudo-label supervision signals through physical rules, avoiding overfitting noise in the student model during training.

[0034] Step S103: Dynamically assign Bayesian weights to the verified data to obtain weighted data with confidence labels.

[0035] Preferably, step S103 further includes: Step S1031: Construct a Bayesian prior distribution based on the historical data quality statistical characteristics, and generate an initial confidence benchmark value for each formatted data. Step S1032: Perform a physical constraint consistency check on the verified data, calculate the likelihood probability that each data point conforms to the transient physical model of the equipment, and form quantitative credible evidence. Step S1033: Combining the Bayesian prior distribution and credible evidence, the posterior probability is obtained through Bayesian update and used as a dynamic weight to generate weighted data with credibility labels.

[0036] It is worth noting that the likelihood probability calculation is based on the deviation measure of the transient physical model of the equipment. Specifically: First, the theoretical physical response curve of the equipment is established, such as the pressure drop-flow characteristic curve of a steam pipeline network. Then, the waveform of the actual collected data is matched with the theoretical curve, such as calculating the Euclidean distance or Pearson correlation coefficient. The smaller the deviation, the higher the likelihood probability. Combining the prior probability formed by the historical accuracy of the data source, the posterior confidence is calculated using Bayes' formula: P(A|B) = {P(B|A)P(A)} / {P(B)}, such as P(real|data) and P(jump|data). Finally, this posterior probability is mapped to a dynamic weight in the 0-1 interval to guide the subsequent differential weighting of the loss function and the online distillation process.

[0037] This embodiment transforms the "hard decision" of physical constraint cleaning into a probabilistic "soft quantification" by dynamically allocating Bayesian weights, thus forming a data credibility assessment. Instead of simply eliminating or fixing weights, this embodiment differentiates the data validated by the physical model, reducing the weight of corrected "instrument jump" data and assigning even higher weights than normal data to confirmed "real overload" data. This is because these physically verified extreme conditions contain crucial information about the equipment's extreme operating limits, effectively preventing subsequent models from developing cognitive biases regarding boundary conditions in scenarios with few samples, thus realizing the weight logic that "the more realistic the anomaly, the more important it is."

[0038] Step S104: Based on the hysteresis response characteristics of the coupling between the medium characteristics and the equipment dynamics, the weighted data is time-series misaligned to obtain the time-compensated weighted data.

[0039] Preferably, step S104 further includes: Step S1041: Based on the characteristics of the energy medium and the dynamic coupling characteristics of the equipment, extract the lag response time under typical working conditions offline and construct a lag time parameter library labeled with influencing factors. Step S1042: Based on the real-time production instructions and operating parameters, call the lag time parameter library and combine it with the lightweight dynamic adaptation model to dynamically calculate the precise lag time of the current running point and generate a timing misalignment mapping table that clearly defines the energy consumption response time offset of each production instruction. Step S1043: Perform forward or backward time-series shift on the energy consumption response sequence according to the mapping table, align the production instructions with the energy consumption data timestamp phase, verify the alignment effect through mutual information, and then fuse the data after lag compensation with the Bayesian dynamic weights to output the time-compensated weighted data.

[0040] It is worth noting that the coupling characteristics of energy medium and equipment power refer to the interaction between the physical laws of different energy transmission media and the response mechanism of energy-consuming equipment. For example, the thermal inertia of steam causes temperature changes to lag behind valve opening, and the fuel-steam conversion delay in the boiler combustion system further amplifies this lag; the electromagnetic inertia and mechanical rotational inertia of a motor jointly determine its power response curve. The coupling of these two factors determines the end-to-end time delay characteristics from "production command issuance → medium parameter change → equipment energy consumption response". Those skilled in the art can determine the coupling characteristics according to specific energy scenarios, and this embodiment does not impose specific limitations on this.

[0041] In this embodiment, the lag time parameter library for influencing factors can specifically adopt an offline, multi-dimensional lookup table structure. Each record includes, but is not limited to: ① Energy medium type: steam / electricity / gas; ② Equipment model: chain grate boiler / variable frequency motor; ③ Typical operating point: 30% / 50% / 80% load rate; ④ Ambient temperature: -10℃ / 25℃ / 40℃; ⑤ Baseline lag time: steam pressure response lag of 600 seconds; ⑥ Lag time fluctuation range: ±15%. This lag time parameter library provides an initial mapping for online calculations, avoiding cold starts.

[0042] It is also worth noting that production instructions are discrete events in the production scheduling plan, such as blast furnace charging and motor start-up and shutdown, and usually have precise timestamps. Operating parameters are real-time continuous variables, such as current load rate, medium temperature, and pipeline pressure, reflecting the instantaneous state of the equipment. In this embodiment, they are used to superimpose dynamic corrections on the baseline lag time.

[0043] Lightweight dynamic adaptation models can employ multilayer perceptron (MLP) or Takagi-Sugeno fuzzy inference (TS) systems. Both are characterized by their lightweight, high speed, and low computational consumption, with fewer than 10,000 parameters. They can achieve microsecond-level lag time calculations at the edge, accurately fitting the nonlinear relationship between operating conditions and lag without complex feature engineering. In application, the input is a vector of operating condition parameters, and the output is a lag time correction coefficient. For example, when the steam load rate increases from 50% to 80%, the model outputs a correction coefficient of 0.85, adjusting the baseline lag of 600 seconds to 510 seconds to reflect the lag reduction effect caused by the increased medium flow rate at full load.

[0044] To further understand the timing misalignment alignment principle of this embodiment, we provide a specific example of steam heating below: (1) Scene Boiler B-01 in a steel plant supplies steam to rolling mill R-03. The production scheduling instruction is "8:00:00 Rolling mill starts, steam demand suddenly increases from 5t / h to 20t / h". (2) Process description Offline database construction: Based on the characteristics of steam medium (high thermal inertia, pipeline capacity delay) and boiler-pipeline coupling characteristics (combustion → steam generation → transmission), the baseline lag time of 600 seconds for boiler B-01 at 80% load is entered into the parameter database, and influencing factors such as load rate and steam temperature are associated.

[0045] Online adaptation: Real-time operating conditions are 85% load rate and 180℃ steam temperature. The lightweight MLP model outputs a correction coefficient of 0.95. The current lag time is dynamically calculated to be 600 × 0.95 = 570 seconds, generating the following: Figure 2 The timing misalignment mapping table shown clearly indicates that the 08:00 instruction corresponds to the response start time of 08:09:30.

[0046] Read the offset -570 seconds from the mapping table, perform batch timestamp subtraction on the original data, and the underlying layer can generate a time series compensation detail table as shown in Table 1. Then, perform a shift operation to generate a table as shown in Table 1. Figure 3 The timestamp phases are aligned as shown. After completion, the mutual information value between "08:00 instruction" and "08:00:00 traffic change" is calculated to be 0.82 (>0.6 threshold), verifying that the alignment is effective; Table 1. Timing Compensation Details

[0047] exist Figure 3In response to the "mill start" command issued at 08:00:00 on 202X-XX-XX, the system calls the lag parameter library and, combined with real-time operating conditions (load rate 85%), calculates the precise lag time of the steam pipeline network to be 570 seconds. During compensation, all original data timestamps are subtracted by 570 seconds: the steady-state data (flow rate 5t / h) before the commands at 08:00:00 and 08:05:00 are shifted to 07:50:30 and 07:55:30 respectively as response baselines; the first response point (flow rate 6t / h) at 08:09:30 is aligned to the command time at 08:00:00; and subsequent data at 08:10:00 (10t / h), 08:15: The 15t / h at 08:00, 19t / h at 08:20:00, and 20t / h at 08:25:00 were sequentially shifted to 08:00:30, 08:05:30, 08:10:30, and 08:15:30, respectively, thus fully preserving the ramp-up transition process of "5t / h→6t / h→10t / h→15t / h→19t / h→20t / h" and achieving causal phase alignment between production instructions and energy consumption response.

[0048] It is important to note that the time-compensated weighted data uses Bayesian posterior probability to achieve dynamic credibility labeling. The weight corresponding to each timestamp is not a static label, but a real-time score that integrates the historical reliability (prior) of the data source and the current physical model validation results (likelihood). See Table 1 and... Figure 3 As shown, the time-compensated weighted data uses Bayesian dynamic weighting to quantify the credibility of each point: the steady-state data before the instruction (07:50:30-07:55:30) lacks physical event support and inherits the historical prior weight of 0.7 from the data source; the first response point at 08:00:00 is verified by the transient model as the start of a real overload, with a likelihood probability of 0.85, and the weight jumps to 0.85; as the flow rate increases to 10t / h and 15t / h, the consistency between the data and the equipment dynamic coupling model continues to improve, and the weight dynamically increases to 0.9-0.95, indicating that this period is a high-quality extreme working condition sample. This weight sequence is not statically labeled but continuously updated online—if sensor drift causes a decrease in prior knowledge or an increase in the error of the teacher model, the weight of the corresponding period will be adaptively adjusted upward, forcing the student model to focus on learning.

[0049] Step S105: Extract 3% of the total data from the time-compensated weighted data to train the DistilBERT teacher model.

[0050] Preferably, step S105 further includes: Step S1051: Perform stratified sampling on the time-compensated weighted data according to energy scenarios and feature distribution to obtain 3% of the total data volume of teacher training samples. Step S1052: Perform time series normalization and standardization preprocessing on the teacher training samples, and then divide them into training set, validation set and test set. More specifically, the ratio of training set, validation set and test set is 7:2:1. Step S1053: Using the initial DistilBERT teacher model, the trained DistilBERT teacher model is obtained by sequentially training it in stages on the training set, tuning its parameters on the validation set, and evaluating it on the test set.

[0051] Understandably, the stratified sampling process employed in this embodiment first divides the data according to energy scenarios, including key areas such as load forecasting and equipment fault diagnosis. Then, within each scenario, it further subdivides according to data characteristic distribution, such as normal / abnormal states and peak / off-peak periods. This stratified sampling strategy ensures that the samples not only cover different application scenarios but also reflect the characteristic changes of data under different states, thereby improving the model's generalization ability and adaptability. In this way, the training samples can more realistically represent the data diversity in real-world applications, providing more comprehensive support for model training.

[0052] As a preferred approach, the sampling range is 1%-5% of the total data volume. This range ensures sufficient coverage of typical patterns such as seasons, peaks and valleys, and holidays, without increasing the burden of annotation and processing due to excessive samples.

[0053] It can also be understood that the aforementioned time series normalization refers to transforming time series data from different devices and systems, with different sampling frequencies and lengths, into a fixed-length time series to facilitate batch processing by the model. Specific methods include using a sliding window to extract high-frequency data, padding or interpolating to expand low-frequency data, and segmenting and length aligning non-time series text data. This embodiment does not specifically limit these methods.

[0054] The aforementioned standardization preprocessing refers to eliminating numerical differences between different energy media and different magnitudes of characteristics, enabling the model to focus on relative changes in the data rather than absolute values. Specific methods include Z-score standardization and Min-Max normalization, scaling the data to a uniform numerical range. This embodiment does not impose specific limitations on this.

[0055] It is important to note that, compared to the unification of structure, time sequence, and semantics, the time sequence regularization and standardization preprocessing here is a technical processing of the data to ensure the consistency of the input data in terms of time scale and feature magnitude, while the unification of structure, time sequence, and semantics is the overall standardization and specification of the data, covering multiple dimensions such as character encoding, time format, and semantic measurement.

[0056] In this embodiment, the teacher training samples are divided into a 70% training set, a 20% validation set, and a 10% test set. The training set is used for the main training process of the model, providing sufficient data to ensure that the model can fully learn the patterns and features in the data. Therefore, a larger training set helps to improve the model's fitting ability and stability. The validation set is used for hyperparameter tuning and model evaluation during the training process. By evaluating the model's performance on the validation set, hyperparameters such as the learning rate and regularization parameters can be adjusted in a timely manner to prevent the model from overfitting. The test set is used for the final evaluation after the model training and tuning are completed. It provides a completely independent dataset to evaluate the model's generalization ability in practical applications, and its results can reflect the model's performance on new data.

[0057] It is worth noting that the teacher model selected in this embodiment is DistilBERT because of its lightweight architecture and strong semantic extraction capabilities. Compared with the complete BERT, DistilBERT reduces the number of parameters by 40% while retaining 95% of the performance, which is suitable for the computing power efficiency requirements of the teacher deployment. By transferring and loading pre-trained weights from the energy domain, it can quickly capture static semantic features such as device topology and long-cycle patterns, providing high-confidence soft-label supervision signals for scenarios with scarce annotations, driving the student LSTM to accurately learn dynamic temporal patterns, and achieving an optimized trade-off between teacher scale and distillation efficiency.

[0058] In step S106, the DistilBERT teacher model generates pseudo-labels from the formatted data to drive the online distillation training of the LSTM student model, and updates the teacher parameters based on the actual running error.

[0059] Preferably, step S106 further includes: Step S1061: Extract unlabeled samples covering all working conditions from the uniformly formatted data, construct a training corpus according to the DistilBERT input specification, and simultaneously prepare the corresponding time-series preprocessed LSTM input format. Step S1062: Input the training corpus into the DistilBERT teacher model to generate high-confidence pseudo-labels. The initial LSTM student model learns the teacher output synchronously. The knowledge transfer intensity is controlled by the dynamic temperature coefficient. Online distillation is completed to obtain the trained LSTM student model. Step S1063: Deploy the trained LSTM student model to the real scene, compare its predicted output with the real data in real time, and calculate the prediction error. Step S1064: Based on the cumulative error, the DistilBERT teacher model parameters are updated online using an exponential moving average method.

[0060] As is understandable, the construction of the training corpus and LSTM input format in this embodiment involves, more specifically: first, preprocessing the raw data by cleaning, deduplication, and normalization to improve data quality; then, extracting relevant features, such as temporal and statistical features, according to task requirements; next, dividing the data into training, validation, and test sets; and finally, converting the temporal data into a three-dimensional tensor input format acceptable to LSTM to ensure that the data can be effectively processed by the model. This process is readily understood by those skilled in the art and will not be elaborated upon further here.

[0061] It's also understandable that high-confidence pseudo-labels represent results with high confidence when the model predicts unlabeled data. Expanding the scale of labeled data can reduce labeling costs and improve the model's generalization ability and performance. Training with these pseudo-labels can effectively compensate for insufficient labeled data and improve the overall performance of the model.

[0062] This example uses the Exponential Moving Average (EMA), a method for smoothing parameter updates. By weighting the current and historical parameter values, it avoids drastic fluctuations in parameter updates, improves model stability, and adapts to data changes. Each update calculates new parameter values ​​based on the prediction error and historical parameters, where the weights of the historical parameters are determined by the exponential decay coefficient, thus achieving smooth updates and continuous model optimization.

[0063] It is worth noting that LSTM was chosen in this embodiment due to its strong ability to embed physical constraints and its efficiency in edge processing. Specifically, by simplifying the gating mechanism, physical laws such as power ramping and energy conservation can be naturally injected. Furthermore, its O(n) complexity and millions of parameters are well-suited to the computing power of energy nodes, and its gating unit has a better response speed to transient fluctuations. At the same time, the teacher DistilBERT slowly updates the teacher parameters through EMA based on the cumulative error gradient of the student LSTM, ensuring that the teacher's knowledge is not fixed and can continuously adapt to equipment aging and operating condition drift.

[0064] In step S107, during the inference phase, the trained DistilBERT teacher model extracts static semantic features, and the trained LSTM student model extracts dynamic temporal features. After unidimensional attention weighted fusion, the prediction result is output.

[0065] Preferably, step S107 further includes: In step S1071, during inference, the trained DistilBERT teacher model extracts static semantic features, and the trained LSTM student model extracts dynamic temporal features. After parameterless alignment, the two features are dynamically weighted and fused using a one-dimensional attention coefficient α to form fused features. The one-dimensional attention coefficient α is optimized by the LSTM student model during distillation training. Step S1072: Input the fused features into the downstream task head and output the final result; Step S1073: Before deployment, perform overall pruning, FP16 quantization and export of the teacher-student-task head fusion as a whole, and perform real-time inference at the edge.

[0066] It is understood that the parameterless alignment used in this embodiment refers to aligning features from different sources directly without adjusting their dimensions or scale through additional parameters or network layers during feature fusion. Specifically, before fusing the static semantic features extracted by the DistilBERT teacher model and the dynamic temporal features extracted by the LSTM student model, it is necessary to ensure their consistency in dimension and scale. This can be achieved through operations such as broadcasting or reshaping. This allows the two types of features to be directly fused without introducing additional computational overhead, making feature fusion concise and efficient, and avoiding increasing model complexity.

[0067] It's also understandable that the one-dimensional attention coefficient α is a weighting coefficient used for feature fusion, optimized by the LSTM student model during distillation training. Specifically, during training, the LSTM student model learns how to better combine the static semantic features of the DistilBERT teacher model with its own dynamic temporal features to optimize the attention coefficient α. The value of α ranges from 0 to 1, dynamically adjusting the weight ratio of the two features in the fusion process. In this way, the model can adaptively determine the respective proportions of static semantic features and dynamic temporal features in the final feature representation according to specific task requirements. This dynamic weighted fusion mechanism allows the model to more flexibly respond to changes in feature importance under different scenarios, thereby improving the model's generalization ability and performance.

[0068] Meanwhile, the resulting fused feature is a comprehensive feature representation. First, parameter-free alignment ensures consistency in dimensionality and scale between the two features. Then, a weighted sum of the two features is obtained using a one-dimensional attention coefficient α. This fused feature integrates static semantic information and dynamic temporal information, enabling a more comprehensive description of the input data's features. The formation of the fused feature allows the model to fully utilize feature information from different sources, improving its understanding and processing capabilities for complex data, thereby enhancing the performance of downstream tasks.

[0069] In this embodiment, the downstream task head refers to a network layer or module in the model used to process specific downstream tasks. After feature fusion is completed, the fused features are input into the downstream task head for further processing and computation, ultimately outputting the task result. The design of the downstream task head depends on the specific task type, for example: Classification tasks: Fully connected layers or softmax layers can be used to output class probabilities; Regression tasks: Linear regression layers can be used to output continuous value predictions; Sequence generation task: Recurrent neural networks (RNNs) or Transformer decoders can be used to generate sequence outputs.

[0070] Downstream task heads can transform fused features into task-specific output formats, enabling the model to complete various downstream tasks. By flexibly designing downstream task heads, the same feature fusion model can be adapted to multiple application scenarios, improving the model's versatility and practicality.

[0071] It is worth noting that the teacher, DistilBERT, handles static semantics (device topology, long-cycle patterns) and provides global supervision, while the student, LSTM, focuses on dynamic temporal sequences (power fluctuations, transient impacts) to achieve millisecond-level responses. The two are dynamically weighted through single-dimensional attention α, decoupling long short-term memory and avoiding the single model from getting caught in global-local optimization conflicts. This achieves optimized resource-precision configuration, meaning that the teacher is responsible for high-precision knowledge distillation and participates in the output, while the student is responsible for lightweight real-time inference.

[0072] This embodiment adopts a semi-supervised pre-training and online distillation training framework, which requires only a small number of labeled samples. It continuously utilizes unlabeled data in combination with actual running error feedback, effectively alleviating the problem of label scarcity and forming a closed loop of data self-repair and model self-optimization.

[0073] Example 2

[0074] like Figure 4 As shown, in one embodiment, a semi-supervised training device for a trusted AI model for energy heterogeneous data compensation is provided. This device includes: The Unified Formatting Module 2001 is used to unify the structure, semantics, and temporal sequence of scattered raw data to obtain formatted data. The Physical Constraint Cleaning Module 2002 is used to perform physical rule verification on outliers in formatted data based on the equipment's rated parameters and transient characteristics. The data quality weighting module 2003 is used to dynamically assign Bayesian weights to the verified data to obtain weighted data with confidence labels. The timing lag compensation module 2004 is used to perform timing misalignment alignment on the weighted data based on the lag response characteristics of the medium characteristics and the dynamic coupling of the equipment, so as to obtain the timing-compensated weighted data. The teacher model generation module 2005 is used to extract a small number of samples from the time-compensated weighted data to train the DistilBERT teacher model. The online distillation training module 2006 is used to generate pseudo-labels from formatted data to drive the online distillation training of the LSTM student model from the DistilBERT teacher model, and to update the teacher parameters based on the actual running error. The Inference Prediction Module 2007 is used in the inference stage. The trained DistilBERT teacher model extracts static semantic features, and the trained LSTM student model extracts dynamic temporal features. The prediction results are then output after being weighted and fused by unidimensional attention.

[0075] Below, we provide some comparative experiments to further illustrate this embodiment, as follows: To verify the effectiveness of this technical solution in energy data scenarios, we conducted comparative experiments on two typical tasks—provincial power grid load forecasting and photovoltaic power plant fault diagnosis—based on the semi-supervised training framework constructed in the embodiments. The experiments followed the data governance process (physical constraint cleaning → Bayesian weighting → lag compensation) and distillation training mechanism (teacher DistilBERT generating pseudo-labels → student LSTM online learning → one-dimensional attention fusion) of the embodiments, and were tested in a unified edge deployment environment.

[0076] 1. Construction of the experimental dataset This experiment used two real energy datasets, with strict control over the annotation ratio to simulate the real bottleneck of "low annotation, high real-time" in the industry. The power load data came from the peak and valley records of a provincial power grid in 202X, and the photovoltaic fault data came from the equipment monitoring logs of a 50,000 kW power plant. Both datasets underwent the unified time-series formatting and quality verification of the implementation example. The basic information of the experimental datasets is shown in Table 2. Table 2. Basic Information of the Experimental Dataset

[0077] 2. Experimental Variables and Control Group Design Three control groups were set up, all deployed on the same edge computing node to ensure environmental consistency. Group 1 simulated the traditional fully supervised mode, Group 2 adopted the general semi-supervised method, and Group 3 adopted the complete technical solution of this implementation. The experimental variables and control group configurations are shown in Table 3. Table 3. Experimental variables and control group configuration

[0078] 3. Evaluation Indicator System For the energy scenario, the evaluation indicators cover three dimensions: prediction accuracy, deployment efficiency, and model credibility. In addition to the conventional mean absolute error (MAE) and macro-average F1 score, intermediate process indicators such as pseudo-label accuracy and teacher-student prediction consistency are added to verify the effectiveness of the semi-supervised mechanism. The experimental evaluation indicator system is shown in Table 4. Table 4. Test Evaluation Index System

[0079] 4. Load Prediction Test Results The load forecasting test results are shown in Table 5. Based on 80,000 power load data points, this embodiment achieved superior performance compared to the control group with only 3% labeled samples: the mean absolute error (MAE) decreased to 45.2 kW, the root mean square error (RMSE) decreased to 56.8 kW, the inference time was 48 ms / sample, the model parameters were reduced to 9.8 M, and the pseudo-label accuracy reached 89%. Compared to the fully supervised group and the general semi-supervised group, this embodiment improved inference efficiency by at least 2.4 times, achieving dual optimization of accuracy and real-time performance. Table 5. Load Prediction Test Results

[0080] 5. Fault Diagnosis Test Results The results of the fault diagnosis test are shown in Table 6. On 40,000 photovoltaic fault data, this embodiment achieved a recall rate of 83.7% for extreme conditions such as short circuit and open circuit, and reduced the false negative rate to 4.5%. This shows that the physical constraint cleaning and Bayesian weighting mechanism effectively ensured the credibility of extreme samples and avoided noise amplification in semi-supervised learning.

[0081] Table 6. Fault Diagnosis Test Results

[0082] 6. Conclusion ① Energy data labeling is costly. In this embodiment, with only 3% labeling, the utilization rate of unlabeled data is increased by 67% through pseudo-label screening with a confidence level >0.85 and online distillation, effectively making up for the deficiency of insufficient labeling. ② The parameterless alignment + single-dimensional attention design reduces the amount of fusion computation by 75%, the inference time is reduced from 168ms to 48ms, and the model size is compressed to 5.2MB, which is fully compatible with the computing power constraints of Intel i5 at the edge. ③ Compared with full supervision, the load prediction MAE decreased by 33.3%, the fault diagnosis F1 value increased by 14.8%, and the extreme condition recall rate increased by 21.6%, indicating that physical constraint cleaning and Bayesian weighting effectively utilize real overload samples.

[0083] This embodiment improves the utilization rate of unlabeled data through pseudo-label screening under low labeling and online distillation, achieves lightweight edge deployment through parameterless alignment and one-dimensional attention, and improves prediction accuracy and recall for extreme conditions through physical constraint cleaning and Bayesian weighting, thus realizing a more efficient and real-time energy AI.

[0084] Example 3

[0085] exist Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 5 As shown, the electronic device may include a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute the semi-supervised training method for the energy heterogeneous data trusted compensation AI model as described in Embodiment 1.

[0086] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0088] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0089] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0090] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, Ruby, and Go, as well as conventional procedural programming languages ​​such as "C" or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0091] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A semi-supervised training method for a reliable compensation AI model for heterogeneous energy data, characterized in that, The method includes the following steps: The scattered raw data is unified in terms of structure, semantics, and temporal sequence to obtain formatted data; Physical rule verification is performed on outliers in formatted data based on the equipment's rated parameters and transient characteristics. Dynamically assign Bayesian weights to the validated data to obtain weighted data with confidence labels; Based on the hysteresis response characteristics of the coupling between the medium properties and the equipment dynamics, the weighted data is time-series misaligned to obtain time-compensated weighted data. A small number of samples were extracted from the time-compensated weighted data to train the DistilBERT teacher model; The DistilBERT teacher model generates pseudo-labels from formatted data to drive online distillation training of the LSTM student model, and updates the teacher parameters based on the accumulated actual running error. During the inference phase, the trained DistilBERT teacher model extracts static semantic features, and the trained LSTM student model extracts dynamic temporal features. The prediction results are then output after being weighted and fused by unidimensional attention.

2. The semi-supervised training method for the AI ​​model of reliable compensation for heterogeneous energy data according to claim 1, characterized in that, The step of unifying the structure, semantics, and temporal sequence of the dispersed raw data to obtain formatted data further includes: The scattered raw data is transformed into formatted data with unified structure, time series, and semantics. The structure encoding rule is UTF-8; the time series is unified with the timestamp of year-month-day-hour-minute-second; and the semantics are unified with energy consumption / kilowatt-hour.

3. The semi-supervised training method for the AI ​​model of reliable compensation for heterogeneous energy data according to claim 1 or 2, characterized in that, The step of performing physical rule verification on outliers in formatted data based on equipment rated parameters and transient characteristics also includes: The system verifies the formatted data for abnormal values ​​based on the equipment's rated parameters. If the values ​​exceed the limits, it calls the equipment's transient physical model to distinguish between instrument fluctuations and actual overloads. Fluctuation data is corrected by reverse calculation using the equipment's transient physical model, while overload data is approved after being marked with an overload flag. If the values ​​do not exceed the limits, the system approves directly.

4. The semi-supervised training method for the AI ​​model of reliable compensation for heterogeneous energy data according to claim 1 or 2, characterized in that, The step of dynamically assigning Bayesian weights to the verified data to obtain weighted data with confidence labels further includes: A Bayesian prior distribution is constructed based on the statistical characteristics of historical data quality, and an initial confidence baseline value is generated for each formatted data. Perform a physical constraint consistency check on the verified data, calculate the likelihood probability that each data point conforms to the transient physical model of the equipment, and form quantitative credible evidence. By combining the Bayesian prior distribution with credible evidence, the posterior probability is obtained through Bayesian update and used as a dynamic weight to generate weighted data with credibility labels.

5. The semi-supervised training method for the AI ​​model of reliable compensation for heterogeneous energy data according to claim 1 or 2, characterized in that, The step of aligning the weighted data according to the hysteresis response characteristics of the coupling between the medium properties and the equipment dynamics to obtain the time-compensated weighted data further includes: Based on the characteristics of energy medium and equipment dynamic coupling characteristics, the lag response time under typical working conditions is extracted offline, and a lag time parameter library labeled with influencing factors is constructed. Based on real-time production instructions and operating parameters, the system calls the lag time parameter library and combines it with a lightweight dynamic adaptation model to dynamically calculate the precise lag time of the current running point and generate a timing misalignment mapping table that clearly defines the energy consumption response time offset for each production instruction. The energy consumption response sequence is shifted forward or backward according to the mapping table to align the production instructions with the energy consumption data timestamp phase. After verifying the alignment effect through mutual information, the data after lag compensation is fused with Bayesian dynamic weights to output the time-compensated weighted data.

6. The semi-supervised training method for the AI ​​model of reliable compensation for heterogeneous energy data according to claim 1 or 2, characterized in that, The step of extracting a small number of samples from the time-compensated weighted data to train the DistilBERT teacher model also includes: The time-compensated weighted data was stratified and sampled according to energy scenarios and feature distributions to obtain a small number of teacher training samples. The teacher training samples are subjected to time series normalization and standardization preprocessing, and then divided into training set, validation set and test set; The initial DistilBERT teacher model is used to obtain the trained DistilBERT teacher model by sequentially training it in stages on the training set, tuning its parameters on the validation set, and evaluating it on the test set.

7. The semi-supervised training method for the AI ​​model of reliable compensation for heterogeneous energy data according to claim 1 or 2, characterized in that, The steps described above, including generating pseudo-labels from formatted data to drive online distillation training of the LSTM student model using the DistilBERT teacher model and updating teacher parameters based on accumulated actual running errors, also include: Unlabeled samples covering all working conditions are extracted from formatted data, and a training corpus is built according to the DistilBERT input specification. Simultaneously, the corresponding time-series preprocessed LSTM input format is prepared. The training corpus is input into the DistilBERT teacher model to generate high-confidence pseudo-labels. The initial LSTM student model learns the teacher output synchronously. The knowledge transfer intensity is controlled by a dynamic temperature coefficient, and online distillation is completed to obtain the trained LSTM student model. The trained LSTM student model is deployed to a real-world scenario, and its predicted output is compared with the real data in real time to calculate the prediction error. Based on the cumulative error, the DistilBERT teacher model parameters are updated online using an exponential moving average.

8. The semi-supervised training method for the AI ​​model of reliable compensation for heterogeneous energy data according to claim 1 or 2, characterized in that, The step of extracting static semantic features from the trained DistilBERT teacher model and extracting dynamic temporal features from the trained LSTM student model during the inference phase, and then outputting the prediction result after unidimensional attention weighted fusion, also includes: During inference, the trained DistilBERT teacher model extracts static semantic features, and the trained LSTM student model extracts dynamic temporal features. After parameterless alignment, the two features are dynamically weighted and fused using a one-dimensional attention coefficient α to form fused features. The one-dimensional attention coefficient α is optimized by the LSTM student model during distillation training. The fused features are input into the downstream task head, and the final result is output. Before deployment, the teacher-student-task header fusion is pruned, FP16 quantized, and exported in ONNX format for real-time inference at the edge.

9. A semi-supervised training device for a reliable compensation AI model for heterogeneous energy data, characterized in that, The device includes: The unified formatting module is used to unify the structure, semantics, and temporal sequence of scattered raw data to obtain formatted data; The physical constraint cleaning module is used to perform physical rule verification on outliers in formatted data based on the equipment's rated parameters and transient characteristics. The data quality weighting module is used to dynamically assign Bayesian weights to the verified data to obtain weighted data with confidence labels. The timing lag compensation module is used to perform timing misalignment alignment on the weighted data based on the lag response characteristics of the medium characteristics and the dynamic coupling of the equipment, so as to obtain the timing-compensated weighted data. The teacher model generation module is used to extract a small number of samples from the time-compensated weighted data to train the DistilBERT teacher model. The online distillation training module is used to generate pseudo-labels from formatted data to drive the online distillation training of the LSTM student model from the DistilBERT teacher model, and to update the teacher parameters based on the actual running error. The inference prediction module is used in the inference stage. The trained DistilBERT teacher model extracts static semantic features, and the trained LSTM student model extracts dynamic temporal features. The prediction results are then output after being weighted and fused by unidimensional attention.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the semi-supervised training method for the AI ​​model for reliable compensation of heterogeneous energy data as described in any one of claims 1 to 8.

Images