Power transformer multivariate sequence prediction method and system based on behavior learning enhancement
By constructing a behavioral learning network and combining it with a joint loss function of oil temperature and load consistency penalty terms, the problem of oil temperature continuity and dynamic consistency in multivariate sequence prediction of power transformers is solved, thereby improving the reliability and adaptability of the prediction.
Patent Information
- Application Number
- CN202611379281.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-07
- Publication Date
- 2026-11-06
AI Technical Summary
Existing multivariate sequence prediction methods for power transformers cannot guarantee the temporal continuity and thermal inertia of oil temperature prediction results, the dynamic consistency between load changes and oil temperature changes, and lack a unified model enhancement mechanism, resulting in prediction results that do not conform to the operating laws of transformers.
A behavior-learning-based reinforcement method is adopted. By constructing a behavior-learning network, a joint loss function is built by combining the oil temperature prediction utility term, the boundary penalty term, and the load-oil temperature dynamic consistency penalty term. This is used to train a multivariate prediction model for power transformers, ensuring that the prediction results conform to the thermal operation law of transformers.
It improves the reliability of multi-step oil temperature prediction and the accuracy of over-temperature warning, reduces prediction results that do not conform to the operating rules of transformers, and enhances the adaptability and pluggability of the model.
Smart Images
Figure CN123050755A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multivariate prediction technology for power transformers, specifically to a method and system for multivariate sequence prediction of power transformers based on behavior learning enhancement. Background Technology
[0002] Power transformers are crucial electrical equipment in power transmission and distribution systems. During transformer operation, changes in load levels cause variations in their thermal state, and transformer oil temperature reflects the equipment's thermal operating condition. Predicting transformer oil temperature over a future period provides data for equipment operation status monitoring, over-temperature early warning, and operation and maintenance.
[0003] The oil temperature variation of a power transformer is typically related to historical oil temperatures and external electrical loads, exhibiting a certain degree of thermal inertia, temporal continuity, and periodicity. Taking power transformer temperature data as an example, its historical operating data can include useful load at high load levels, unused load at high load levels, useful load at medium load levels, unused load at medium load levels, useful load at low load levels, unused load at low load levels, and transformer oil temperature. These different load variables collectively reflect the external load state of the transformer under different operating conditions, while oil temperature is used to characterize the thermal state of the transformer.
[0004] Existing multivariate sequence prediction methods for power transformers can employ linear prediction models, self-attention mechanism-based prediction models, periodic prediction models, or multi-scale prediction models to generate future oil temperature predictions based on historical load and historical oil temperature. Current methods typically use point-level prediction errors such as mean square error and mean absolute error as model training targets to reduce the numerical deviation between predicted and actual oil temperatures.
[0005] However, training a multivariate sequence prediction model for power transformers using only point-level prediction errors has at least the following problems.
[0006] First, point-level prediction errors mainly constrain the numerical deviation at a single prediction moment, making it difficult to guarantee that the complete oil temperature prediction curve satisfies the temporal continuity and thermal inertia of transformer oil temperature changes. Especially within long prediction intervals, prediction results that do not conform to the thermal response law of transformers may occur, such as sudden changes in oil temperature and local abnormal oscillations.
[0007] Second, existing training objectives do not explicitly evaluate the dynamic consistency between load changes and oil temperature changes. Predictive models may produce future oil temperature curves with small numerical errors but that do not match historical load conditions. For example, they may predict a large jump in oil temperature when the load change is small, or predict an unreasonable drop in oil temperature when the load continues to increase.
[0008] Third, existing methods typically do not explicitly incorporate operational constraints such as the allowable oil temperature range and the rate of oil temperature change during training. As the prediction time increases, the model output may exceed the preset oil temperature range, or the temperature change between adjacent prediction times may exceed the allowable range, affecting the reliability of the prediction results for transformer thermal condition monitoring and over-temperature early warning.
[0009] Fourth, existing improvement methods mainly enhance prediction accuracy by changing the network structure of the oil temperature prediction model. When it is necessary to adapt to linear prediction networks, self-attention-based prediction networks, periodic prediction networks, or multi-scale prediction networks, it is usually necessary to modify the model structure separately. There is a lack of a training mechanism that can uniformly enhance different oil temperature prediction models without changing the main structure of the prediction backbone network.
[0010] Therefore, a technical solution for multivariate prediction of power transformers is needed. Based on numerical prediction using historical load and historical oil temperature, the solution further evaluates the oil temperature boundary, temperature rise change, and load-oil temperature dynamic consistency of the multivariate prediction sequence of candidate power transformers. The evaluation results are then incorporated into the model training process to reduce prediction results that do not conform to the operating laws of transformers and improve the reliability of multi-step oil temperature prediction and over-temperature early warning. Summary of the Invention
[0011] To address the problem that existing multivariate prediction models for power transformers mainly rely on point-level prediction errors, making it difficult to guarantee that the predicted oil temperature sequence meets the oil temperature safety boundary, temperature rise continuity, and load-oil temperature dynamic consistency, this invention provides a multivariate sequence prediction method and system for power transformers based on behavior learning enhancement.
[0012] This invention generates a multivariate prediction sequence for future power transformers based on historical external load characteristics and historical oil temperatures. It then calculates the operational rationality score of this prediction sequence using a behavioral learning network. The operational rationality score is determined at least based on an oil temperature prediction utility term, an oil temperature boundary penalty term, and a load-oil temperature dynamic consistency penalty term. During training, the point-level oil temperature prediction loss is combined with the operational rationality score to form a joint training objective. This reduces numerical prediction errors in the oil temperature prediction model while minimizing prediction results that do not conform to the thermal operation characteristics of power transformers.
[0013] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for multivariate sequence prediction of power transformers based on behavior learning enhancement, comprising: The historical multivariate sequence of power transformers is acquired and preprocessed as input. The historical multivariate sequence of power transformers includes the external power load characteristics under at least one load level and the transformer oil temperature. A differentiable baseline prediction model was used to generate multivariate prediction sequences for power transformers. A behavior learning network is constructed to receive historical multivariate sequences and multivariate prediction sequences of power transformers and output behavior utility scores. The behavior utility scores are calculated based on learnable utility terms, constraint penalty terms, and consistency penalty terms. Calculate the point-level prediction loss between the multivariate predicted sequence of power transformers and the actual multivariate sequence of power transformers; A joint loss is constructed based on a weighted combination of point-level prediction loss and behavioral utility score. During training, this joint loss is minimized to simultaneously reduce prediction error and maximize the behavioral utility of the multivariate prediction sequence of power transformers. The parameters of the baseline model and the behavioral learning network are jointly updated end-to-end. After training, the baseline model was used to predict multivariate sequences of power transformers.
[0014] In one embodiment, the variables in the historical multivariate sequence of the power transformer include: useful load at the high load level, useless load at the high load level, useful load at the medium load level, useless load at the medium load level, useful load at the low load level, useless load at the low load level, and transformer oil temperature.
[0015] In one embodiment, the constructed behavior learning network is used to receive historical multivariate sequences and predicted multivariate sequences of power transformers and output behavior utility scores. The behavior utility scores are calculated based on learnable utility terms, constraint penalty terms, and consistency penalty terms, specifically including: Constructing conditional behavior utility functions Calculate the behavioral utility score: ; in, For the historical multivariate series of power transformers, For multivariate prediction sequences of power transformers, For learnable potential utility functions, For learnable inequality constraint functions, For learnable equality constraint functions, , , These are non-negative weighting coefficients. For the tanh function, For the ReLU function, It is an absolute value function. As a learnable utility term, To constrain penalties, This is a consistency penalty item.
[0016] In one embodiment, calculating the point-level prediction loss between the multivariate predicted sequence of the power transformer and the actual multivariate sequence of the power transformer specifically includes: The point-level prediction loss is a weighted combination of one or more of the following: mean squared error, mean absolute error, Huber loss, quantile loss, and probability prediction negative log-likelihood loss. ; Point-level prediction loss when using mean square error for: ; in, For the corresponding real future sequence, express Norm.
[0017] In one embodiment, the construction of the joint loss based on the weighted combination of point-level prediction loss and behavioral utility score specifically includes: Joint losses for: ; in, To adjust the weights between numerical fitting and behavioral consistency, For point-level loss prediction, For multivariate prediction sequences of power transformers, To predict the multivariate sequence of power transformers Historical multivariate series of power transformers Input behavior learning network The output is the behavioral utility score.
[0018] In one embodiment, the differentiable baseline prediction model is selected from any one of linear prediction models, Transformer prediction models, multi-scale hybrid models, periodic modeling models, and frequency domain modeling models.
[0019] In one embodiment, the process of using a baseline model for multivariate sequence prediction of power transformers after training specifically includes: During the inference phase, only the trained baseline prediction model is retained to output the multivariate prediction sequence for power transformers.
[0020] In a second aspect, the present invention provides a multivariate prediction system for power transformers based on behavior learning enhancement, for implementing the method of any embodiment of the first aspect, comprising: The data processing module is used to acquire and preprocess historical multivariate sequences of power transformers; The baseline prediction module uses a differentiable baseline prediction model to generate multivariate prediction sequences for power transformers. The behavior learning module is used to construct a behavior learning network and output the behavior utility score under historical conditions. The joint training module is used to calculate point-level prediction loss, construct joint loss, and perform end-to-end joint parameter updates. The prediction output module is used to output the multivariate prediction sequence of the power transformer after training is completed; The model management module is used to configure the baseline prediction model type, prediction length, input length, and weight parameters in the joint loss.
[0021] The system and method in this invention correspond to each other; the specific technical solutions applicable to the method are also applicable to the system.
[0022] Compared with the prior art, the beneficial technical effects of the present invention are: (1) It provides behavioral structure supervision in addition to prediction error. Traditional mean squared error (MSE) or mean absolute error (MAE) mainly constrains point-by-point distance. This invention further evaluates the utility, constraint satisfaction and variable consistency of the complete future sequence under given historical conditions through behavioral learning network, thereby enhancing the overall rationality of prediction output.
[0023] (2) It has strong pluggability. This invention does not require the reconstruction of the main structure of the baseline prediction model. It only needs to connect the behavior learning network and joint loss during the training phase to adapt to different prediction backbone networks such as linear models, Transformer models, multi-scale models, periodic models and frequency domain models.
[0024] (3) It can provide learnable output space gradients for the prediction model. The behavioral utility term can transmit the gradient back to the baseline prediction model through the prediction sequence, which helps the model learn potential dynamic laws and behavioral constraints beyond simple point-level errors, and reduces the problem of unreasonable overall trends caused by local numerical fitting.
[0025] (4) Experimental results show that this method has a general enhancement effect under multiple model, multiple dataset, and multiple prediction length conditions. Specifically, on the ETTh1, ETTh2, ETTm1, and ETTm2 datasets, models such as SparseTSF, DLinear, Informer, Autoformer, SOFTS, Crossformer, TimeMosaic, WPMixer, TimeMixer, and CycleNet were compared using mean squared error (MSE) and mean absolute error (MAE) as evaluation metrics. After behavior learning enhancement, the error decreased on most prediction lengths and dataset combinations. Specific results can be found in Tables 2 and 3. The above results indicate that behavior learning can serve as an effective supplement to traditional prediction loss.
[0026] (5) It has application extension value. Since the input of the behavioral learning network is historical sequence and prediction sequence, the present invention can be transferred to different multivariate dynamic systems such as energy, transportation, meteorology, public health, disease trends, financial markets and industrial equipment status; in the inference stage, the behavioral utility score can also be used as the basis for prediction credibility, anomaly detection or candidate scheme screening. Attached Figure Description
[0027] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a diagram of the architecture of the present invention. Detailed Implementation
[0028] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.
[0029] like Figure 1 As shown, a method for multivariate sequence prediction of power transformers based on behavior learning enhancement in this invention includes the following steps: S1. Obtain and preprocess the historical multivariate sequence of the power transformer as input. The historical multivariate sequence of the power transformer includes the external power load characteristics under at least one load level and the transformer oil temperature. S2, using a differentiable baseline prediction model to generate multivariate prediction sequences for power transformers; S3, construct a behavior learning network to receive historical multivariate sequences and multivariate prediction sequences of power transformers and output behavior utility scores. The utility scores are calculated based on learnable utility terms, constraint penalty terms, and consistency penalty terms. S4, calculate the point-level prediction loss between the multivariate prediction sequence of the power transformer and the multivariate true sequence of the power transformer; S5 constructs a joint loss based on a weighted combination of point-level prediction loss and behavioral utility score. During training, this joint loss is minimized to simultaneously reduce the multivariate error of the power transformer and maximize behavioral utility. The parameters of the baseline model and the behavioral learning network are jointly updated end-to-end. S6. After training, the baseline model is used to predict multivariate sequences of power transformers.
[0030] The following is a breakdown into several parts.
[0031] 1. Acquire and preprocess multivariate time series data.
[0032] Obtain historical multivariate time series samples from the target dynamic system, denoted as the historical multivariate series of the power transformer. The multivariable true sequence of power transformers is ,in Indicates the input length. Indicates the predicted length. This represents the variable dimension. Preprocessing is performed on the original sequence, including missing value handling, outlier filtering, timestamp sorting, variable normalization or standardization, and sliding window segmentation, to obtain the training sample set. .
[0033] The variables in the historical multivariate sequence of power transformers include: useful load at high load level, useless load at high load level, useful load at medium load level, useless load at medium load level, useful load at low load level, useless load at low load level, and transformer oil temperature.
[0034] High load, medium load, and low load are load levels classified according to the transformer load factor (i.e., the ratio of actual load to rated capacity). In one embodiment, the classification can be as follows: a load factor greater than 40% is defined as a high load level, a load factor between 20% and 40% is defined as a medium load level, and a load factor less than 20% is defined as a low load level. Useful load refers to the effective load portion that supports the normal operation of the transformer, while useless load refers to the ineffective load portion caused by factors such as harmonics and reactive power.
[0035] 2. Construct a baseline prediction model.
[0036] Choose any differentiable time series prediction model as the baseline prediction model. This includes, but is not limited to, linear prediction models, Transformer prediction models, multi-scale hybrid models, periodic modeling models, or frequency domain modeling models. The baseline prediction model generates multivariate prediction sequences for power transformers based on historical multivariate sequences. ; Specifically, the baseline prediction model can be selected from Dlinear model, SparseTSF model, Informer model, Autoformer model, SOFTS model, Crossformer model, TimeMixer model, WPMixer model, CycleNet model or TimeMosaic model depending on the application scenario.
[0037] 3. Construct a behavioral learning network.
[0038] Building behavioral learning networks Used to receive historical multivariate sequences of power transformers and multivariate prediction sequences for power transformers The algorithm outputs the behavioral utility score of the multivariate prediction sequence of the power transformer under given historical conditions. The behavioral learning network includes a utility term, a constraint penalty term, and a consistency penalty term, and its differentiable behavioral utility function is expressed as: .
[0039] in, The learnable potential utility function is the utility term for the rationality of oil temperature prediction operation; The learnable inequality constraint function is the oil temperature boundary and temperature rise rate penalty term; The learnable equality constraint function is the load-oil temperature dynamic consistency penalty term; , , These are non-negative weighting coefficients; , , It can be a differentiable mapping function or a neural network submodule.
[0040] In the specific implementation, As a behavioral utility module. Historical multivariate sequence of power transformers. and power transformer multivariate prediction series The input can be directly into the network, or it can be flattened, convolutionally encoded, attention-encoded, or pooled before being input into the network to obtain the behavioral utility score. .
[0041] 4. Calculate the traditional prediction loss.
[0042] For the baseline prediction model output With power transformer multivariable real sequence Calculate point-level prediction loss Preferably, mean squared error is used: .
[0043] In other implementations, a weighted combination of mean absolute error (MAE), Huber loss, quantile loss, negative log-likelihood of probability prediction, or multiple losses may also be used.
[0044] 5. Calculate the learning utility of behavioral patterns.
[0045] The multivariate prediction sequence of power transformers output by the baseline prediction model Historical multivariate series of power transformers Inputting the data into a behavior learning network yields a behavior utility score: .
[0046] The higher the score, the more the predicted sequence matches the underlying behavioral structure learned by the behavioral learning network; the lower the score, the more the predicted result may be close to the true value locally, but there are unreasonable aspects in terms of overall sequence behavior, variable coupling, or system constraints.
[0047] 6. Construct joint training objectives.
[0048] Combining the traditional prediction loss with the behavioral learning utility term yields the following joint loss: ,in .
[0049] Due to minimization during training The behavioral learning term is preceded by a negative sign, so the objective is equivalent to maximizing the behavioral utility of the multivariate prediction sequence of power transformers while reducing prediction error. Used to adjust the weights between numerical fitting and behavioral consistency; when When it degenerates to the original baseline prediction model training; when Behavioral learning enhancement is enabled at that time.
[0050] 7. End-to-end joint optimization.
[0051] In each training batch, the baseline prediction model is first used. Generate multivariate prediction sequences for power transformers Then, the behavior learning network Output behavior utility Subsequently, based on the joint losses Simultaneously update the parameters of the baseline prediction model. Parameters of behavioral learning networks The behavioral learning term exerts a gradient effect on the baseline prediction model parameters through the predicted output, and its gradient form is: .
[0052] The second term uses the chain rule to pass the gradient of the behavioral utility network with respect to the predicted output to the baseline prediction model, thereby guiding the prediction results to move towards the output region of higher behavioral utility.
[0053] 8. Forecasting and Deployment.
[0054] After training is complete, only the trained baseline prediction model can be retained during the inference phase. It can make rapid predictions while simultaneously preserving the behavioral learning network. The system scores, rearranges, detects anomalies, or assesses the credibility of candidate predictions. It also receives new historical multivariate sequences of power transformers. Output power transformer multivariate prediction sequence It can also output behavioral utility scores, variable contribution information, or risk warnings.
[0055] This invention constructs a conditional behavior utility function to evaluate the rationality of the behavior of a multivariate prediction sequence of a power transformer under given historical multivariate series of power transformers. The behavior utility function is jointly optimized end-to-end with the baseline prediction model, maximizing the behavior utility of the multivariate prediction sequence of the power transformer while reducing numerical errors. A pluggable enhancement framework is formed, enabling the behavior learning module to adapt to different time series prediction backbone models such as DLinear, Informer, Autoformer, SparseTSF, SOFTS, Crossformer, TimeMixer, WPMixer, CycleNet, and TimeMosaic.
[0056] Figure 2 Displaying historical multivariate sequences of power transformers Input baseline prediction model Generate multivariate prediction sequences for power transformers ,at the same time and Input behavior learning network Output behavior utility It is then used in conjunction with traditional prediction loss to form a joint loss for training.
[0057] Table 1 shows the training pseudocode. This diagram includes window sample construction, baseline prediction model prediction, prediction loss calculation, behavior utility calculation, joint loss construction, and backpropagation update.
[0058] Table 1
[0059] Tables 2 and 3 are schematic diagrams comparing the experimental results. These diagrams are used to show the changes in mean squared error (MSE) and mean absolute error (MAE) before and after introducing the behavior learning term on multiple public time series datasets and multiple prediction backbone models.
[0060] Table 2
[0061] Table 3
[0062] Example 1: Behavioral learning-enhanced prediction based on MLP-type models.
[0063] Taking DLinear as an example, the historical input length The predicted length can be set to 96, 192, 336, or 720. It can be set to 96, 192, 336, or 720. The input data is... Multivariate series. First, standardization is performed on the raw data, and a sliding window sample is constructed. DLinear is used as the baseline prediction model. Input historical multivariate sequence of power transformers Post-output power transformer multivariate prediction sequence Constructing behavioral learning networks The network receives and The concatenated representation, after passing through a multilayer perceptron or other differentiable encoder, outputs a scalar behavioral utility. This is calculated during training. , then calculate Ultimately Update DLinear and After training, use DLinear to output future multivariate prediction results.
[0064] Example 2: Behavioral learning-enhanced prediction based on Transformer-type models.
[0065] In this embodiment, the baseline prediction model can be Informer, Autoformer, or Crossformer. These models internally employ attention mechanisms, sequence decomposition mechanisms, or cross-variable dependency modeling mechanisms to generate future sequences. The behavior learning network does not depend on the internal structure of the baseline prediction model; it only receives historical inputs and predicted outputs, and therefore can be directly integrated into the training process. For each batch, the Transformer-type model outputs... Subsequently, the behavioral learning network calculates the behavioral utility at the overall sequence level; the joint loss maximizes the behavioral utility while optimizing numerical errors. This embodiment is suitable for load, traffic, or weather forecasting tasks with strong variable correlations and significant long-term dependencies.
[0066] Example 3: Behavioral learning-enhanced prediction based on periodic or multi-scale models.
[0067] In this embodiment, the baseline prediction model can be selected from TimeMixer, WPMixer, CycleNet, TimeMosaic, or SparseTSF. These models utilize multi-scale mixing, frequency decomposition, periodic representation, or sparse periodic structures, respectively. The behavior learning network complements these structures: the baseline prediction model generates multivariate prediction sequences for power transformers with corresponding inductive biases, while the behavior learning network evaluates whether these multivariate prediction sequences satisfy the underlying behavior structure under historical conditions. After training, it is possible to deploy only the baseline prediction model to reduce inference overhead, or to deploy the behavior learning network simultaneously for prediction scoring.
[0068] Example 4: Training Algorithm.
[0069] Input training data Baseline prediction model Behavioral learning networks Weight And the training epoch. For each training epoch, iterate through the mini-batch (X,Y) and execute: A1: ; A2: ; A3: ; A4: ; A5: Update simultaneously via backpropagation and ; A6: Output the trained prediction model until the validation set error or the number of training rounds meets the stopping condition.
[0070] Example 5: Deployment and Application.
[0071] In energy load scenarios, It can include historical load, temperature, holiday variables, and equipment status variables. For the future load curve, For predicting future load curves; behavioral utility can be learned from electricity consumption behavior, weather influences, and periodic constraints. In transportation scenarios, It can include multi-segment traffic flow, speed, and congestion indicators, and behavioral utility can be learned from road network coupling and commuting cycles. In public health or disease trend scenarios, It can include historical case counts, testing volume, environmental factors, or medical resource indicators, and behavioral utility can be learned about transmission mechanisms and intervention constraints.
[0072] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.
[0073] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0074] Based on the description of the above method embodiments, the present invention also provides a system. The system may be a system that uses software (applications), modules, components, servers, clients, etc., using the methods described in the embodiments of this specification, combined with necessary implementation hardware. Since the implementation schemes and methods for solving the problem are similar, the specific system implementations in the embodiments of this specification can be found in the implementations of the foregoing methods. Although the system is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.
[0075] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0076] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.
[0077] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for multivariate sequence prediction of power transformers based on behavior learning reinforcement, characterized in that, include: The historical multivariate sequence of power transformers is acquired and preprocessed as input. The historical multivariate sequence of power transformers includes the external power load characteristics under at least one load level and the transformer oil temperature. A differentiable baseline prediction model was used to generate multivariate prediction sequences for power transformers. A behavior learning network is constructed to receive historical multivariate sequences and multivariate prediction sequences of power transformers and output behavior utility scores. The behavior utility scores are calculated based on learnable utility terms, constraint penalty terms, and consistency penalty terms. Calculate the point-level prediction loss between the multivariate predicted sequence of power transformers and the actual multivariate sequence of power transformers; A joint loss is constructed based on a weighted combination of point-level prediction loss and behavioral utility score. During training, this joint loss is minimized to simultaneously reduce prediction error and maximize the behavioral utility of the multivariate prediction sequence of power transformers. The parameters of the baseline model and the behavioral learning network are jointly updated end-to-end. After training, the baseline model was used to predict multivariate sequences of power transformers.
2. The method for multivariate sequence prediction of power transformers based on behavior learning enhancement according to claim 1, characterized in that, The variables in the historical multivariate sequence of the power transformer include: useful load at the high load level, useless load at the high load level, useful load at the medium load level, useless load at the medium load level, useful load at the low load level, useless load at the low load level, and transformer oil temperature.
3. The method for multivariate sequence prediction of power transformers based on behavior learning enhancement according to claim 1, characterized in that, The constructed behavior learning network is used to receive historical multivariate sequences and predicted multivariate sequences of power transformers and output behavior utility scores. The behavior utility scores are calculated based on learnable utility terms, constraint penalty terms, and consistency penalty terms, specifically including: Constructing conditional behavior utility functions Calculate the behavioral utility score: ； in, For the historical multivariate sequence of power transformers, For multivariate prediction sequences of power transformers, For learnable potential utility functions, For learnable inequality constraint functions, For learnable equality constraint functions, , , These are non-negative weighting coefficients. For the tanh function, For the ReLU function, It is an absolute value function. As a learnable utility term, To constrain penalties, This is a consistency penalty item.
4. The method for multivariate sequence prediction of power transformers based on behavior learning enhancement according to claim 1, characterized in that, The calculation of the point-level prediction loss between the multivariate predicted sequence and the actual multivariate sequence of the power transformer specifically includes: The point-level prediction loss is a weighted combination of one or more of the following: mean squared error, mean absolute error, Huber loss, quantile loss, and probability prediction negative log-likelihood loss. ; Point-level prediction loss when using mean square error for: ； in, For the corresponding real future sequence, express Norm.
5. The method for multivariate sequence prediction of power transformers based on behavior learning enhancement according to claim 1, characterized in that, The construction of the joint loss based on a weighted combination of point-level prediction loss and behavioral utility score specifically includes: Joint losses for: ； in, To adjust the weights between numerical fitting and behavioral consistency, For point-level loss prediction, For multivariate prediction sequences of power transformers, To predict the multivariate sequence of power transformers Historical multivariate series of power transformers Input behavior learning network The output is the behavioral utility score.
6. The method for multivariate sequence prediction of power transformers based on behavior learning enhancement according to claim 1, characterized in that, The differentiable baseline prediction model is selected from any one of the following: linear prediction model, Transformer prediction model, multi-scale hybrid model, periodic modeling model, and frequency domain modeling model.
7. The method for multivariate sequence prediction of power transformers based on behavior learning enhancement according to claim 1, characterized in that, After training, the baseline model is used to predict multivariate sequences of power transformers, specifically including: During the inference phase, only the trained baseline prediction model is retained to output the multivariate prediction sequence for power transformers.
8. A multivariate prediction system for power transformers based on behavior learning reinforcement, used to implement the method as described in any one of claims 1 to 7, characterized in that, include: The data processing module is used to acquire and preprocess historical multivariate sequences of power transformers; The baseline prediction module uses a differentiable baseline prediction model to generate multivariate prediction sequences for power transformers. The behavior learning module is used to construct a behavior learning network and output the behavior utility score under historical conditions. The joint training module is used to calculate point-level prediction loss, construct joint loss, and perform end-to-end joint parameter updates. The prediction output module is used to output the multivariate prediction sequence of the power transformer after training is completed; The model management module is used to configure the baseline prediction model type, prediction length, input length, and weight parameters in the joint loss.