Substation-level demand response potential estimation method based on physical constraint and self-supervised representation learning

Abstract

The invention discloses a transformer substation level demand response potential estimation method based on physical constraint and self-supervised representation learning, and relates to the field of power demand response, and the method comprises the steps: S1, generating a TCL prior clue through combining the cooperative decomposition and physical consistency selection of two channels; s2, learning to obtain the discriminative TCL time sequence representation of the TCL priori clue; s3, performing prediction based on discriminative TCL time sequence representation to obtain an original predicted value, embedding a physical security layer behind a prediction head, and analyzing and projecting the obtained original predicted value to a time-varying convex polyhedron in real time to obtain an optimized predicted value meeting operation and physical boundaries; and S4, generating a peak clipping amount delta P (t) and a transferable energy envelope for a scheduling instruction based on the optimized predicted value. According to the method, high precision and high robustness are kept under heterogeneous working conditions such as commercial high fluctuation and industrial stability through transformer substation level TCL prediction and DR adjustable margin estimation.

Description

Technical Field

[0001] This application relates to the field of power demand response, and in particular to a method for estimating substation-level demand response potential based on physical constraints and self-supervised representation learning. Background Technology

[0002] Demand response (DR), as the most cost-effective and efficient flexibility resource, has been widely validated to reduce peak loads by 5%–15% and significantly improve grid resilience at the system level. However, in actual distribution network operations, substations, as the first information-physical aggregation node for distributed power sources, producers, consumers, and concentrated loads, typically only obtain SCADA aggregated measurements with low time resolution (5–15 minutes), lacking terminal sub-metering and fine-grained energy consumption tags. Constrained by factors such as privacy compliance, high transformation costs, and data ownership disputes, the large-scale deployment of advanced metering infrastructure (AMI) and smart meters in existing substations has progressed slowly, with the vast majority of operating sites facing a long-term data-poor situation of "only having a master meter, but no sub-meters."

[0003] Under the aforementioned practical constraints, accurately quantifying the adjustable margin of demand response at the substation level has significant engineering value: On the one hand, it is necessary to ensure that the DR commitment does not trigger local thermal constraints or voltage over-limit. On the other hand, it is necessary to identify the spatial heterogeneity and flexibility caused by different end uses such as residential thermal loads (TCL), electric vehicle charging, and small commercial equipment, so as to provide a reliable basis for day-ahead and intraday scheduling, capacity market pricing, and ancillary service procurement.

[0004] However, relying solely on SCADA total measurements lacking detailed information constitutes a fundamental obstacle to quantitatively assessing DR capabilities at relevant operational nodes. Traditional statistical or empirical models rapidly degrade in generalization ability and stability under conditions of complex load composition, strong meteorological coupling, and nonlinear superposition of user behavior. Summary of the Invention

[0005] The purpose of this application is to provide a substation-level demand response potential estimation method based on physical constraints and self-supervised representation learning. This method comprehensively utilizes aggregated load and meteorological information, while introducing physical consistency constraints and unsupervised / self-supervised representation learning to ensure that the prediction results satisfy the electrical and thermal physical boundaries in both the training and inference stages. This enables accurate quantification and practical application of substation-level demand response potential under the premise of privacy protection and low retrofit cost.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for estimating substation-level demand response potential based on physical constraints and self-supervised representation learning, the method comprising: S1. The acquired substation total load and synchronous temperature data are preprocessed using dual-channel collaborative decomposition and physical consistency selection to generate TCL prior clues. ; S2. Obtaining TCL Prior Clues Based on Mask Reconstruction and Temporal Comparison Learning Discriminative TCL timing characterization; S3. Based on the discriminative TCL timing representation, the original predicted value is obtained. A physical safety layer is embedded after the prediction header. Combining the physical safety layer and the original predicted value, an optimized predicted value that satisfies the operational and physical boundaries is obtained. ; S4. Generate the peak reduction amount ΔP(t) and the transferable energy envelope EY for scheduling instructions based on the optimized predicted values.

[0007] Furthermore, the generation of TCL prior cues and multi-scale components in S1 also includes: Based on the substation-level demand response scenario, empirical mode decomposition is used to obtain the intrinsic mode and the remainder term: ; ; in, This represents the total load of the substation. To synchronize temperature or equivalent external disturbance; The remaining load at the current time point; This represents the temperature remainder at the current time point. This represents the load-side intrinsic mode at the current time point; The intrinsic temperature mode at the current time point; Based on time-varying delay correlation, each load mode With temperature mode group Perform a correlation scan and record the maximum correlation across time lags. for: ; Temperature coefficient; Set threshold Combining the maximum correlation across time delays With threshold The temperature-driven candidate mode set is obtained as follows: ; k For load-side intrinsic modes The index; set S represents the set of load mode numbers determined to be temperature-driven; Introducing lumped parameter RC thermal balance proxy constraints, the indoor equivalent temperature is calculated recursively using external air temperature and TCL equivalent power. Indoor equivalent temperature satisfy: ; Outdoor ambient temperature (which can be obtained from a weather station or numerical weather forecast) is aligned with the total load time of the substation. This is the equivalent effect of temperature control load on indoor heat balance. The equivalent power of the temperature-controlled load is obtained by superimposing the candidate load modes driven by temperature. The equivalent heat transfer / cooling (heating) efficiency coefficient is used to map electrical power to equivalent heat flux. This is an equivalent internal thermal disturbance term used to characterize the impact of non-temperature-controlled heat sources such as personnel activities, equipment heat dissipation, and solar radiation on the indoor thermal state. In implementation, it can be set as a constant, a slowly varying term, or estimated from historical data. Indoor equivalent temperature Discretized as: ; Where, a is the indoor thermal inertia coefficient, used to characterize the degree of self-maintenance of the current indoor equivalent temperature to the next moment; b is the outdoor temperature conduction coefficient, used to characterize the influence weight of the outdoor ambient temperature on the indoor thermal state; c is the equivalent effect coefficient of the temperature control load, used to map the power of the temperature control load to the change in indoor temperature; d is the equivalent constant disturbance term, used to comprehensively characterize the influence of internal heat sources, unmodeled thermal disturbances and model discretization errors. The obtained multi-scale clues Superimposed to generate TCL prior clues .

[0008] Furthermore, in S2, prior clues for TCL are obtained based on mask reconstruction and timing comparison. Discriminative TCL timing characterization also includes: Robust feature representations of learning time-series data are reconstructed using masks, and the ability of the model to distinguish between similar and different segments in the time dimension is enhanced by using a time-series comparison mechanism. Mask reconstruction includes: reconstruction using mean square error: ; Time-series comparisons include: InfoNCE-style contrast loss to enhance discriminative power and noise resistance. ; Where sim() represents the cosine similarity. For temperature coefficient, Positive samples for homologous enhancement fragments; Combination and The weighted sum training of the encoder and decoder; Using intraday peaks and troughs as natural positive and negative sample pairs, and combining masked sequence autoencoder pre-training with NT-Xent contrastive loss, the peak-trough contrastive learning framework guides the model to distinguish between high duty cycle and low duty cycle of temperature control load without manual annotation. Based on high-duty-load and low-duty-load operation states, a comparative learning mechanism is used to obtain time-series representations of latent variables such as TCL duty cycle, start-stop dynamics, and meteorological sensitivity.

[0009] Furthermore, in step S3, the original predicted value is obtained by predicting based on the discriminative TCL time series representation. .

[0010] ; Where h is the prediction step index, representing the time step since the current moment. The first step forward One predicted time step, To predict the length of the time domain; Based on the discriminative TCL timing representation, at time... The given future TCL power prediction values ​​at each time step.

[0011] Furthermore, in S3, an optimized prediction value that satisfies the operational and physical boundaries is obtained by combining the physical security layer and the original prediction value. It also includes: Let the prediction time domain be Time step ; Define the operational and thermophysical boundaries: ;

[0012] The safety layer uses quadratic programming to transform the original predicted values Project to feasible region: ; In end-to-end training, strong constraints are approximated using differentiable penalties: ; in , To predict the first in the time domain TCL power prediction values ​​at each time step; This represents the maximum available polymerization power for TCL at the corresponding time step. To predict the time step; As of the date The upper limit of the cumulative available energy at each time step is used to characterize thermal inertia and comfort constraints; The maximum allowable ramp rate for TCL power change between adjacent time steps; These are the penalty weight coefficients for different physical constraints, used to balance the impact of each constraint during the training process.

[0013] Furthermore, combining the root mean square error loss value Compare the loss values and physical constraint loss Construct the comprehensive loss L: ; By combining the comprehensive loss L, we obtain the optimal prediction value that satisfies the operational and physical boundaries. .

[0014] The advantage of using the differentiable penalty approximation strong constraint technique here is that: In the end-to-end training phase, direct solution of non-differentiable projection operators is avoided, allowing physical constraints to participate in gradient backpropagation in a continuous and differentiable form, thereby ensuring that model parameters can be stably optimized through backpropagation. Simultaneously, analytical projection is still used in the inference phase, strictly mapping the prediction results to the time-varying convex polyhedron to ensure that the final output 100% satisfies the operational and thermophysical boundaries. This design balances training differentiability with operational feasibility.

[0015] Furthermore, in step S4, a peak-shaving amount for scheduling instructions is generated based on the optimized predicted value. In addition to the transferable energy envelope EY, it also includes: ; The predicted prior power sequence for TCL; The optimized prediction result is projected through the physical security layer; peak reduction In the first Reduceable power is achieved through scheduling at each prediction time step; Transferable energy envelope EY: EY = .

[0016] According to the substation-level demand response potential estimation method based on physical constraints and self-supervised representation learning provided in this application, this application has the following technical advantages compared with the prior art: (1) Dual-channel collaborative decomposition and physical consistency decomposition technology: For the first time, the load sequence and temperature sequence are simultaneously decomposed by ICEEMDAN, and the dual physical screening criteria of time-varying intrinsic correlation (TDIC) + lumped RC thermal balance mutual information are introduced. Only the IMF component that simultaneously satisfies the strong temperature coupling and the first-order thermal balance equation is retained, thus eliminating the mode mixing and non-physical decomposition caused by traditional single-channel decomposition in principle. This technology does not rely on any labels and can be directly extended to all aggregate prediction scenarios with temperature-driven flexible loads (such as district heating load, industrial electric boiler group, etc.).

[0017] (2) Based on mask reconstruction and time series comparison, on completely unlabeled SCADA historical data, a peak-valley comparison learning framework is first created, which uses intraday peak and valley periods as natural positive and negative sample pairs, and combines mask sequence autoencode pre-training and NT-Xent comparison loss. This enables the model to autonomously learn latent variable representations such as TCL duty cycle, start-stop dynamics, and meteorological sensitivity, and completely get rid of the dependence on real TCL labels. This self-supervised mechanism can be directly transferred to other unlabeled aggregated load decomposition tasks (such as electric vehicle charging load, data center load, etc.).

[0018] (3) Constraint projection technology with differentiable physical safety layer: It is the first to embed a parameter-free and differentiable physical safety layer after the Transformer prediction head, and project the original prediction value into a time-varying convex polyhedron composed of feeder capacity constraint C1 and rolling thermodynamic energy budget constraint C2 in real time. This ensures that the TCL baseline load and power reduction at each moment of the final output meet the dual hard constraints of network and thermophysics. At the same time, it supports end-to-end gradient training and can extract dual variables as operating margin indicators. This differentiable safety layer has universality and can be directly used to force the addition of physical hard constraints (such as photovoltaic output limit, energy storage SOC boundary, line power flow constraint, etc.) in any power prediction scenario, which greatly expands the protection scope.

[0019] (4) Based on the above three points of integration, the SCADA-only demand response potential envelope direct generation technology is realized for the first time. It can directly output the peak reduction amount ΔP(t) and the transferable energy envelope that simultaneously meet the capacity constraints and thermodynamic constraints from coarse-grained SCADA data and can be directly used for scheduling instructions. Moreover, it does not involve any user-side privacy data throughout the process, truly achieving "zero metering transformation, zero labeling, and zero privacy risk". Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a framework diagram of a substation-level demand response potential estimation method based on physical constraints and self-supervised representation learning in one embodiment of this application; Figure 2 A schematic diagram illustrating the global feature importance of SHAP provided in an embodiment of this application; Figure 3 A comparison chart of calculations and actual TCL values ​​provided for one embodiment of this application; Figure 4 This is a comparison chart of predicted and actual TCL values ​​provided in an embodiment of this application; Figure 5 A schematic diagram of a slice of demand response potential on the coldest day, provided for an embodiment of this application; Figure 6 A schematic diagram of a slice of demand response potential on the hottest day, provided for an embodiment of this application; Figure 7 A schematic diagram of a slice of demand response potential during a summer workday, provided as an embodiment of this application; Figure 8 A slice diagram illustrating the demand response potential during a summer weekend, as provided in one embodiment of this application; Figure 9 A schematic diagram of a slice of demand response potential during a transitional season working day, provided for an embodiment of this application; Figure 10 A schematic diagram of a slice of demand response potential during the weekend of a transitional season, provided as an embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1As shown in the embodiments of this application, a method for estimating substation-level demand response potential based on physical constraints and self-supervised representation learning is disclosed. The method includes: S1. Combining the dual-channel collaborative decomposition and physical consistency selection, the acquired substation total load and synchronous temperature data are preprocessed to generate TCL prior clues. ; S2. Obtaining TCL Prior Clues Based on Mask Reconstruction and Temporal Comparison Learning Discriminative TCL timing characterization; S3. Based on the discriminative TCL timing representation, the original predicted value is obtained. A physical safety layer is embedded after the prediction header. Combining the physical safety layer and the original predicted value, an optimized predicted value that satisfies the operational and physical boundaries is obtained. ; S4. Generate the peak reduction amount ΔP(t) and the transferable energy envelope EY for scheduling instructions based on the optimized predicted values.

[0025] Specifically, in engineering applications, this invention can be implemented according to the illustrated process (see diagram). Figure 1 Method Overall Framework) Implementation: Inputs are the total load of the substation and synchronous meteorological / calendar characteristics; outputs thermodynamically feasible temperature-driven components and TCL priors; learn discriminative representations and generate preliminary predictions under unlabeled conditions; project the predictions into the feasible region to obtain the final results; complete multi-scenario evaluation and interpretable output.

[0026] Through the above steps, a station-level DR potential estimation scheme is formed that can still operate stably under conditions of data scarcity, privacy restrictions, and heterogeneous operating conditions, achieving a balance between prediction accuracy, physical feasibility, and engineering deployability.

[0027] This invention constructs an end-to-end method based on "decomposition-characterization-constraint-inference" under the premise of relying solely on substation-level SCADA aggregated measurement and publicly available meteorological data.

[0028] First, load and meteorological signals are co-decomposed and physical consistency screening is implemented to extract temperature-driven and thermodynamically feasible multi-scale modes. Then, a time encoder is trained with self-supervised / contrastive learning to obtain discriminative representations and output TCL power predictions. Finally, the prediction results are projected into the convex feasible region defined by the operating and thermophysical boundaries through a differentiable safety layer, thereby eliminating infeasible solutions such as negative values ​​and exceeding the upper limit from a mechanism perspective.

[0029] Furthermore, the generation of TCL prior cues and multi-scale components in S1 also includes: Assume the total load of the substation is Synchronous temperature or equivalent external disturbance is An improved empirical mode decomposition (such as ICEEMDAN) is used to obtain its intrinsic mode function and remainder term: ; ; in, This represents the total load of the substation. To synchronize temperature or equivalent external disturbance; The remaining load at the current time point; This represents the temperature remainder at the current time point. This represents the load-side intrinsic mode at the current time point; The intrinsic temperature mode at the current time point; Based on time-varying delay correlation (TDIC) for each load mode Perform a correlation scan with the temperature mode set; Let the maximum correlation across time delays be . ; in, Temperature coefficient; Set threshold Combining the maximum correlation across time delays With threshold The temperature-driven candidate mode set is obtained as follows: ; k For load-side intrinsic modes The index; set S represents the set of load mode numbers determined to be temperature-driven; To avoid pseudo-modes that are "related but not physically feasible," a lumped-parameter RC thermal balance surrogate constraint is introduced. The indoor equivalent temperature is calculated recursively using the external air temperature and the TCL equivalent power. Indoor equivalent temperature satisfy: ; Outdoor ambient temperature (which can be obtained from a weather station or numerical weather forecast) is aligned with the total load time of the substation. This is the equivalent effect of temperature control load on indoor heat balance. The equivalent power of the temperature-controlled load is obtained by superimposing the candidate load modes driven by temperature. The equivalent heat transfer / cooling (heating) efficiency coefficient is used to map electrical power to equivalent heat flux. This is an equivalent internal thermal disturbance term used to characterize the impact of non-temperature-controlled heat sources such as personnel activities, equipment heat dissipation, and solar radiation on the indoor thermal state. In implementation, it can be set as a constant, a slowly varying term, or estimated from historical data. Indoor equivalent temperature Discretized as: ; Where, a is the indoor thermal inertia coefficient, used to characterize the degree of self-maintenance of the current indoor equivalent temperature to the next moment; b is the outdoor temperature conduction coefficient, used to characterize the influence weight of the outdoor ambient temperature on the indoor thermal state; c is the equivalent effect coefficient of the temperature control load, used to map the power of the temperature control load to the change in indoor temperature; d is the equivalent constant disturbance term, used to comprehensively characterize the influence of internal heat sources, unmodeled thermal disturbances and model discretization errors. Only when the candidate modes are superimposed When (4) has a feasible trajectory that satisfies the temperature-comfort zone, the mode is considered thermodynamically feasible and retained; the resulting multi-scale superposition is the TCL prior of the decomposition stage. .

[0030] Specifically, a dual-channel collaborative decomposition and physical consistency screening method is adopted, with aggregated load and air temperature as parallel inputs in time. Through improved empirical mode decomposition and correlation discrimination, non-temperature-driven oscillations are first eliminated by time correlation, and then thermal balance consistency is verified by lumped parameter RC heat flux proxy. Only intrinsic modes that are "confirmed to be temperature-driven and thermodynamically feasible" are retained, and stable TCL prior clues and multi-scale components are output.

[0031] The improved empirical mode decomposition includes: (1) The traditional single-channel load decomposition is extended to a load-temperature dual-channel coordinated ICEEMDAN decomposition; (2) In the modal screening stage, a time-varying delay correlation criterion considering thermal inertia is introduced to avoid misjudgment caused by static correlation; (3) Further, the lumped parameter RC thermal equilibrium model is used as a physical feasibility constraint, and only intrinsic modes that simultaneously satisfy temperature-driven and thermodynamic consistency are retained; (4) Make the decomposition results a stable prior for subsequent self-supervised learning and physical constraint prediction.

[0032] Through the above improvements, the obtained intrinsic modes are not only mathematically reasonable, but also physically interpretable, fundamentally solving the problems of mode aliasing and non-physical interpretation that exist in traditional empirical mode decomposition in demand response applications.

[0033] Optionally, in step S2, TCL prior clues are obtained based on mask reconstruction and temporal comparison learning. Discriminative TCL timing characterization also includes: Constructing input features Includes historical total load, weather, calendar, etc., in length Sliding window forming sample Encoder (such as a contrastive Transformer) output embedding Robust feature representations of learning time-series data are reconstructed using masks, and a time-series comparison mechanism is used to enhance the model's ability to distinguish between similar and different segments in the time dimension; Mask reconstruction includes: reconstruction using mean square error: ; Time-series comparisons include: InfoNCE-style contrast loss to enhance discriminative power and noise resistance. ; Where sim() represents the cosine similarity. For temperature coefficient, Positive samples for homologous enhancement fragments; Homologous random augmentation includes, but is not limited to, generating positive sample pairs without altering the temporal and physical semantics of TCL through methods such as time-dimensional masking, amplitude perturbation, and subsequence pruning. .

[0034] Combination and The weighted sum training of the encoder and decoder; Using intraday peaks and troughs as natural positive and negative sample pairs, and combining masked sequence autoencoder pre-training with NT-Xent contrastive loss, the peak-trough contrastive learning framework guides the model to distinguish between high duty cycle and low duty cycle of temperature control load without manual annotation. Based on high-duty-load and low-duty-load operation states, a comparative learning mechanism is used to obtain time-series representations of latent variables such as TCL duty cycle, start-stop dynamics, and meteorological sensitivity.

[0035] Since TCL exhibits significant periodic start-stop and temperature-driven characteristics on an intraday scale, the aforementioned comparative learning mechanism enables the model to autonomously capture the temporal representation of latent variables such as TCL duty cycle period, start-stop dynamics, and meteorological sensitivity, thereby gradually eliminating the dependence on real TCL labels when substations only have aggregated measurements.

[0036] The original predicted value is obtained by predicting the TCL time series representation based on discriminant analysis in S3. .

[0037] ; Where h is the prediction step index, representing the time step since the current moment. The first step forward One predicted time step, To predict the length of the time domain; Based on the discriminative TCL timing representation, at time... The given future TCL power prediction values ​​at each time step.

[0038] Specifically, the self-supervised / contrastive learning representation learning does not require any manual labels. It utilizes pre-text text tasks such as temporal masking, adjacent segment comparison, and cross-scale consistency to train the Transformer encoder to extract discriminative TCL temporal representations, significantly improving robustness to noise, abrupt changes, and cross-condition migration.

[0039] Using intraday peak and trough periods as natural positive and negative sample pairs, a peak-trough contrastive learning framework combining masked sequence autoencoder pre-training and NT-Xent contrastive loss guides the model to distinguish between high-duty and low-duty operation states of temperature-controlled loads without manual annotation. Since TCL exhibits significant periodic start-stop and temperature-driven characteristics on an intraday scale, the aforementioned contrastive learning mechanism enables the model to autonomously capture the temporal representations of latent variables such as TCL duty cycle period, start-stop dynamics, and meteorological sensitivity. This allows the model to gradually reduce its dependence on real TCL labels, even when substations only have aggregated measurements.

[0040] Optionally, in step S3, by embedding a physical security layer after the prediction head, the original prediction value is analyzed and projected onto a time-varying convex polyhedron in real time to obtain an optimized prediction value that satisfies the operational and physical boundaries. ; To eliminate negative values, exceeding the upper limit, and energy / climbing violations, the prediction time domain is set as follows: Time step Define the operational and thermophysical boundaries: ; ; The safety layer uses quadratic programming to transform the original predicted values Project to feasible region: ; In end-to-end training, strong constraints are approximated using differentiable penalties: ; in , To predict the first in the time domain TCL power prediction values ​​at each time step; This represents the maximum available polymerization power for TCL at the corresponding time step. To predict the time step; As of the date The upper limit of the cumulative available energy at each time step is used to characterize thermal inertia and comfort constraints; The maximum allowable ramp rate for TCL power change between adjacent time steps; These are the penalty weight coefficients for different physical constraints, used to balance the impact of each constraint during the training process.

[0041] Combined with root mean square error loss value Compare the loss values and physical constraint loss Construct the comprehensive loss L: ; By combining the comprehensive loss L, we obtain the optimal prediction value that satisfies the operational and physical boundaries. .

[0042] The advantage of using the differentiable penalty approximation strong constraint technique here is that: In the end-to-end training phase, direct solution of non-differentiable projection operators is avoided, allowing physical constraints to participate in gradient backpropagation in a continuous and differentiable form, thereby ensuring that model parameters can be stably optimized through backpropagation. Simultaneously, analytical projection is still used in the inference phase, strictly mapping the prediction results to the time-varying convex polyhedron to ensure that the final output 100% satisfies the operational and thermophysical boundaries. This design balances training differentiability with operational feasibility.

[0043] Specifically, by embedding a physical safety layer after the prediction head, the original prediction value is analyzed and projected onto the time-varying convex polyhedron in real time to obtain the optimized prediction value that satisfies the operational and physical boundaries. The time-varying convex polyhedron is the feasible region jointly defined by the operational and thermophysical boundaries listed below, that is, the time-varying convex feasible set jointly formed by the upper and lower power limits, the rolling energy budget constraints, and the ramping constraints in the prediction time domain.

[0044] Design a differentiable safety layer. At the prediction end, the model output is projected onto a convex feasible region defined by "substation capacity limit, TCL≤total load, rolling thermal budget and comfort temperature zone", etc., to eliminate negative values ​​and solutions that exceed the limit in an optimizable way during the training and inference phases, ensuring that the results meet the operational and physical boundaries in a mechanistic way.

[0045] Optionally, in step S4, a peak-shaving amount for scheduling instructions is generated based on the optimized predicted value. In addition to the transferable energy envelope EY, it also includes: ; The predicted prior power sequence for TCL; The optimized prediction result is projected through the physical security layer; peak reduction In the first Reduceable power is achieved through scheduling at each prediction time step; Transferable energy envelope EY: EY = .

[0046] It also includes optimizing prediction results based on RMSE and MAPE metrics. The degree of deviation relative to the prior TCL power sequence.

[0047] ; ; This is the root mean square error; The mean absolute percentage error.

[0048] Furthermore, it can combine feature attribution (such as temperature, humidity, holidays, and operating hours) to output interpretable conclusions: by quantifying the contribution of input features such as temperature, humidity, wind speed, holiday type, and operating hours, it identifies key driving factors affecting substation-level TCL load and its potential for reduction, and provides the potential variation patterns and confidence boundaries under different meteorological and calendar scenarios. Based on the above interpretable conclusions, it can provide a basis for demand response quota configuration (such as committable peak shaving capacity, transferable energy scale and duration), time-of-use response strategy formulation (such as peak priority, valley replenishment, and comfort margin settings), and operation and maintenance decisions (such as risk warnings under high temperature / high humidity conditions, dispatch command trigger thresholds, and reserve capacity reservations).

[0049] Specifically, an inference and evaluation process is established. The method supports different time resolutions and historical window settings, enabling unified evaluation of heterogeneous sites such as commercial and industrial sites, and providing explanatory outputs for operation and maintenance (such as meteorological driving intensity and sensitive periods), facilitating the rapid invocation of DR quotas, peak and valley management, and backup configurations.

[0050] Through the aforementioned mathematical mechanism, this invention achieves the following: under conditions without terminal labels and without sub-items, stable priors are provided through physically consistent dual-channel screening; self-supervised / contrastive learning obtains transferable temporal representations across operating conditions; and a differentiable safety layer intrinsically guarantees physical and operational feasibility throughout the entire training and inference process. Therefore, substation-level TCL prediction and DR adjustable margin estimation maintain high accuracy and robustness under heterogeneous operating conditions such as high commercial volatility and stable industrial conditions, while also possessing privacy friendliness and engineering deployability.

[0051] Example 2: Publicly available household-side data (methodological validity verification), also including: A publicly available resident-side dataset (Pecan Street, including equipment-level electricity consumption) was selected, based on continuous annual data, with a uniform time resolution of 15 minutes to ensure consistency with the workflow of this invention. Input features included: historical total load (resident aggregated feeder side), meteorological elements (such as temperature, humidity, wind speed, etc.), and contextual variables such as calendar / business hours. To suppress dimensional differences, Z-score normalization was performed on all inputs using standardized parameters fitted only to the training set, and denormalization was performed during the inference phase to restore the normalized data.

[0052] Figure 2 The colors in the graph represent the magnitude of the feature values, and the horizontal axis represents the impact on the model output (SHAP value). The results show that surface temperature / air temperature contributes the most, confirming the thermal driving characteristics of the temperature-controlled load.

[0053] The process settings also include: A dual-channel decomposition and time-delay correlation discrimination were performed on the "total load-temperature" relationship. A lumped-parameter RC thermal balance approximation was used to screen candidate modes for thermophysical feasibility, retaining only intrinsic modes that are "temperature-driven and thermodynamically feasible" to obtain the TCL prior clue. A self-supervised / contrastive learning temporal encoder (Transformer architecture) was trained under unlabeled conditions, with the objective being a joint loss of mask reconstruction and segment comparison. The historical window was set to 24 hours (96 × 15 min), and multi-step feedforward was used for prediction. A differentiable safety layer was used to project the feasible region of the predicted sequence onto boundaries such as "capacity limit, TCL ≤ total load, rolling thermal budget, and ramping constraints." A physical penalty approximation was used during training, and a strict convex projection pruning was used during inference. The training-test split was performed in an 8:2 chronological order to avoid information leakage.

[0054] Table 1 Comparison of effects of different lengths

[0055] As can be seen from the table, the overall error is lowest when the sequence length is 41. Therefore, 41 steps (1-minute granularity experiment) and 96 steps (15-minute granularity experiment) are used as the optimal memory length for the two types of tasks by default.

[0056] The indicators and results also include: (1) Feature screening gain: After using SHAP analysis to identify and remove redundant variables with low contribution, the error decreased from RMSE 5.56 kW and MAPE 20.29% to RMSE 4.80 kW and MAPE 16.98%, verifying the effectiveness of variable optimization.

[0057] Table 2 Comparison of SHAP feature selection before and after optimization

[0058] (2) Effectiveness of TCL extraction: On the resident data with device-level TCL tags, compared with various decomposition / separation algorithms, ICEEMDAN-TDIC is significantly better than STL, WT, VMD, LSTM and NILM in seasonal and multi-scale scenarios, indicating that "temperature-driven + thermophysical consistency" screening can effectively alleviate modal mixing and avoid non-physical understanding.

[0059] Figure 3 The blue area represents the estimated TCL using this method, while the red area represents the actual TCL after cleaning. The error is smaller during the steady-state period at night, while the fluctuation increases during the midday transition period due to users simultaneously turning the device on / off.

[0060] Table 3 Comparison of different decomposition algorithms for TCL load extraction

[0061] (3) Overall prediction comparison: Under the same leak-free protocol and the same data segmentation, the present invention is significantly better than the mainstream deep and tree models in point prediction; removing contrastive learning or removing differentiable security layers will lead to a decrease in stability and feasibility.

[0062] Table 4 Performance Comparison of Prediction Algorithms

[0063] Figure 4 The solid red line represents the present invention; the purple dashed line and the blue dotted line represent the ablation versions with contrast and constraint removal, respectively; the cyan, green, and yellow lines represent LSTM, MLP, and Random Forest. The proposed model is highly consistent with the measured values ​​in terms of peak and valley times and amplitudes; the ablation versions exhibit valley drift and overshoot exceeding the upper limit, confirming the necessity of contrastive learning and physical projection.

[0064] (4) DR potential quantification: In six typical "weather-calendar" slices (weekdays and weekends in summer / transitional seasons, and the coldest / hottest days of the year), based on the physically feasible envelope generated by the differentiable safety layer, the estimated DR head gap is approximately 3–6% peak reduction, with a daily energy peak shift of 16–26 kWh. This level is sufficient to support the formulation of strategies for capacitor tapping, standby unit start-up and shutdown, and paid DR settlement on the distribution network side, and can serve as a basis for seasonal and time-based bidding.

[0065] Figure 5-10 It provides six typical "weather-calendar" slices (weekdays and weekends in summer / transitional seasons, and the coldest / hottest days of the year); Example 3: Aggregated data from Nanchong substation (Verification of engineering portability and robustness).

[0066] Data sources and construction also include: The load data for multiple feeders in Nanchong City, Sichuan Province, China, was selected and categorized into industrial and commercial loads based on load structure. The raw data consisted of active power curves at a 15-minute granularity for multiple feeders within the station. First, at the feeder level, missing data detection and linear interpolation, abnormal peak (3σ) suppression, and timestamp alignment were performed. Then, at the station level, aggregation and summation were performed to obtain the total load. Simultaneously, elements from the Nanchong meteorological station (air temperature, surface temperature, wind speed, precipitation, soil moisture, and solar radiation) were integrated, and context variables such as calendar / operating hours were constructed to form a multidimensional time series with 14-dimensional input and 1-dimensional target. Commercial stations exhibited a clear operating hour-based linear transition, while industrial stations showed a continuous smooth transition over 24 hours. The effective sample size for the entire year was approximately 35,000 time points / stations. The engineering site only had aggregated SCADA data and lacked actual TCL component curves. Therefore, in this embodiment, the TCL sequence is directly generated by modules 1–2 (dual-channel ICEEMDAN + physical consistency filtering) on ​​Total(t) and meteorological / contextual data, and TCL(t) is forced to be ≤ Total(t) throughout the entire process.

[0067] The process setup and training protocol also include: The data preprocessing and standardization strategies remain consistent with those in Example 1. Training and testing are divided in an 8:2 time ratio, with no information leakage.

[0068] Time-frequency decomposition, hysteresis judgment, and RC thermal balance feasibility screening were performed on the "total load-temperature" dual channels. Only the intrinsic modes that are "temperature driven and thermodynamically feasible" were retained to obtain the site-level TCL prior.

[0069] The temporal encoder (Transformer) is trained using self-supervised / contrastive learning without real labels. The loss is a weighted sum of mask reconstruction and segment comparison. The history window is 96×15min (24 h), and multi-step feedforward prediction is used.

[0070] The output sequence is projected using a differentiable safety layer with "capacity limit, TCL≤Total, rolling thermal budget, and ramping constraint"; physical penalty approximation is used during training, and strict pruning is used during inference to ensure zero physical infeasibility.

[0071] Baselines for comparison: LSTM, MLP, Random Forest, HistGBR, and TCN, as well as ablation methods that remove contrastive learning and differentiable constraint layers. All deep models employ Adam, cosine annealing, and gradient clipping, with a unified early stopping criterion (RMSE on the validation set).

[0072] The indicators and comparison results also include: In commercial substations (high volatility, strong operating period effects), the proposed model achieves an RMSE of 48.36 kW and a MAPE of 5.37%, which is the best among all baselines, as shown in Table 5. Table 5

[0073] The proposed model is also optimal in industrial substations (smooth, near-continuous operation), with an RMSE of 24.13 kW and a MAPE of 3.45%, as shown in Table 6. Table 6

[0074] Robustness and maintainability verification also includes: (1) Ablation analysis: After removing contrastive learning, a systematic drift occurred in the business trough of the commercial station; after removing the differentiable safety layer, a slight tendency to exceed the upper limit appeared in the boundary time period. Both indicate that representation learning and physical projection are responsible for noise suppression / alignment and feasibility guarantee, respectively, and neither can be omitted.

[0075] (2) Indirect consistency check: Since there is no real TCL label, this embodiment cannot provide an "estimate-real" superimposed curve; however, in the annual statistics, the TCL estimate is significantly correlated with the intraday temperature change, the commercial weekend shows a stable "negative concave" effect, and the daily curve of the industrial station is smooth and the variance is significantly lower than that of the commercial station, all of which are consistent with the actual operation logic.

[0076] (3) Zero violations: Under the unified early shutdown and no leakage protocol, the predictions of the entire test set all meet the requirements of TCL(t)≤Total(t), capacity limit, rolling thermal budget and ramp limit, and there are no physically infeasible outputs such as negative load / out of bounds.

[0077] Engineering implications and deployability also include: (1) Portability: Under the real-world constraints of distribution networks with “no AMI sub-items and only SCADA + meteorology”, the framework achieves optimal or parallel optimal performance in two typical load structures (high-fluctuation commercial and smooth industrial), and has the portability to be directly implemented.

[0078] (2) Strategy implications: The peak-valley "zigzag transition" of commercial stations is more suitable for refined intraday DR (windows before and after business hours, midday and before closing); while industrial stations are more suitable for continuous peak shaving and valley filling and stable backup configuration.

[0079] (3) Simplified operation and maintenance: Security layer parameters can be obtained from site archives or online identification. In the self-monitoring phase, unlabeled data from recent months can be used for rolling adaptation to form a continuous deployment link with "low maintenance cost".

[0080] In summary, Example 2 demonstrates the portability, stability, and zero infeasibility of the present invention in real-world utility scenarios: even in the absence of real TCL tags, it can stably extract temperature control load features and complete high-quality predictions by relying on "physically consistent decomposition + self-supervised comparison + differentiable security layer", supporting site-level DR quantification and scheduling execution.

[0081] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0082] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (Read-Only Memory). Memory includes ROM, magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0083] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0084] 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.

[0085] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for estimating substation-level demand response potential based on physical constraints and self-supervised representation learning, characterized in that, The method includes: S1. Combining the dual-channel collaborative decomposition and physical consistency selection, the acquired substation total load and synchronous temperature data are preprocessed to generate TCL prior clues. ; S2. Obtaining TCL Prior Clues Based on Mask Reconstruction and Temporal Comparison Learning Discriminative TCL timing characterization; S3. Based on the discriminative TCL timing representation, the original predicted value is obtained. A physical safety layer is embedded after the prediction header. Combining the physical safety layer and the original predicted value, an optimized predicted value that satisfies the operational and physical boundaries is obtained. ; S4. Generate the peak reduction amount ΔP(t) and the transferable energy envelope EY for scheduling instructions based on the optimized predicted values.

2. The method for estimating substation-level demand response potential based on physical constraints and self-supervised representation learning according to claim 1, characterized in that, The generation of TCL prior cues and multi-scale components in S1 also includes: Based on the substation-level demand response scenario, empirical mode decomposition is used to obtain the intrinsic mode and the remainder term: ； ； in, This represents the total load of the substation. To synchronize temperature or equivalent external disturbance; The remaining load at the current time point; This represents the temperature remainder at the current time point. This represents the load-side intrinsic mode at the current time point; The intrinsic temperature mode at the current time point; Based on time-varying delay correlation, each load mode With temperature mode group Perform a correlation scan and record the maximum correlation across time lags. for: ； Temperature coefficient; Set threshold Combining the maximum correlation across time delays With threshold The temperature-driven candidate mode set is obtained as follows: ； k For load-side intrinsic modes The index; set S represents the set of load mode numbers determined to be temperature-driven; Introducing lumped parameter RC thermal balance proxy constraints, the indoor equivalent temperature is calculated recursively using external air temperature and TCL equivalent power. Indoor equivalent temperature satisfy: ； Outdoor ambient temperature aligned with the total load time of the substation; This is the equivalent effect of temperature control load on indoor heat balance. The equivalent power of the temperature-controlled load is obtained by superimposing the candidate load modes driven by temperature. The equivalent heat transfer / cooling efficiency coefficient; This is the equivalent internal thermal disturbance term; Indoor equivalent temperature Discretized as: ； Where a is the indoor thermal inertia coefficient; b is the outdoor temperature conduction coefficient; c is the equivalent effect coefficient of temperature control load; and d is the equivalent constant disturbance term. The obtained multi-scale clues Superimposed to generate TCL prior clues .

3. The method for estimating substation-level demand response potential based on physical constraints and self-supervised representation learning according to claim 1, characterized in that, In S2, prior cues for TCL are obtained based on mask reconstruction and temporal comparison learning. Discriminative TCL timing characterization also includes: Robust feature representations of learning time-series data are reconstructed using masks, and the ability of the model to distinguish between similar and different segments in the time dimension is enhanced by using a time-series comparison mechanism. Mask reconstruction includes: reconstruction using mean square error: ； Time-series comparisons include: InfoNCE-style contrast loss to enhance discriminative power and noise resistance. ； Where sim() represents the cosine similarity. For temperature coefficient, Positive samples for homologous enhancement fragments; Combination and The weighted sum training of the encoder and decoder; Using intraday peaks and troughs as natural positive and negative sample pairs, and combining masked sequence autoencoder pre-training with NT-Xent contrastive loss, the peak-trough contrastive learning framework guides the model to distinguish between high duty cycle and low duty cycle of temperature control load without manual annotation. Based on high-duty-load and low-duty-load operating states, a discriminative TCL timing representation is obtained through a contrastive learning mechanism.

4. The method for estimating substation-level demand response potential based on physical constraints and self-supervised representation learning according to claim 1, characterized in that, The original predicted value is obtained by predicting the TCL time series representation based on discriminant analysis in S3. . ； Where h is the prediction step index, representing the time step since the current moment. The first step forward One predicted time step, To predict the length of the time domain; Based on the discriminative TCL timing representation, at time... The given future TCL power prediction values ​​at each time step.

5. The method for estimating substation-level demand response potential based on physical constraints and self-supervised representation learning according to claim 4, characterized in that, In S3, the optimized prediction value that satisfies the operational and physical boundaries is obtained by combining the physical security layer and the original prediction value. It also includes: Let the prediction time domain be Time step ; Define the operational and thermophysical boundaries: ； The safety layer uses quadratic programming to transform the original predicted values Project to feasible region: ； In end-to-end training, strong constraints are approximated using differentiable penalties: ； in , To predict the first in the time domain TCL power prediction values ​​at each time step; This represents the maximum available polymerization power for TCL at the corresponding time step. To predict the time step; As of the date The maximum cumulative available energy per time step; The maximum allowable ramp rate for TCL power change between adjacent time steps; These are the penalty weight coefficients for different physical constraints, used to balance the impact of each constraint during the training process; Combined with root mean square error loss value Compare the loss values and physical constraint loss Construct the comprehensive loss L: ； By combining the comprehensive loss L, we obtain the optimal prediction value that satisfies the operational and physical boundaries. .

6. The method for estimating substation-level demand response potential based on physical constraints and self-supervised representation learning according to claim 1, characterized in that, In step S4, peak shaving amounts for scheduling instructions are generated based on optimized prediction values. In addition to the transferable energy envelope EY, it also includes: ； The predicted prior power sequence for TCL; The optimized prediction result is projected through the physical security layer; peak reduction In the first Reduceable power is achieved through scheduling at each prediction time step; Transferable energy envelope EY, EY = .

Images