Net load forecasting method and system based on hidden variable decoupling and virtual metering

Abstract

The present application belongs to the technical field of power system load forecasting, and particularly relates to a grid-supply net load forecasting method and system based on hidden variable decoupling and virtual metering, which is suitable for the day-ahead load forecasting and power spot transaction decision-making scenarios of provincial / municipal power grid dispatching centers. The method comprises data acquisition alignment and multi-scale physical feature extraction; source-load blind separation driven by hidden variables; reconstruction of historical full-network electricity load benchmark sequence; double-flow collaborative day-ahead forecasting; cross-attention fusion and probability distribution correction. The method does not rely on underlying subdivided physical quantity measurement, realizes distributed photovoltaic virtual accurate metering based on hidden variable modeling, accurately decouples source and load hidden variables, reconstructs grid-supply net load benchmark, and predicts the full-network day-ahead grid-supply net load based on source-load dynamic coupling probability, thereby solving the actual problems of increased random fluctuation of provincial / municipal grid grid-supply net load and difficulty in high-precision prediction under the background of high penetration rate of distributed photovoltaic power generation.

Description

Technical Field

[0001] This invention belongs to the field of power system load forecasting technology, specifically involving a method and system for forecasting net grid load based on latent variable decoupling and virtual metering, applicable to day-ahead load forecasting and electricity spot trading decision-making scenarios in provincial / municipal power grid dispatch centers. Background Technology

[0002] In new power systems, distributed photovoltaic (PV) systems have high penetration rates and are generally self-consumable. Many residential low-voltage grid-connected distributed PV systems typically lack power generation data acquisition devices, making real-time power output impossible to monitor. The net grid load at the provincial / municipal level, calculated by subtracting distributed PV generation from the actual electricity load, is significantly affected by fluctuations in actual electricity load and the randomness of distributed PV generation, making accurate prediction difficult. Source-load decoupling is a crucial prerequisite for net grid load forecasting; the accuracy of decoupling between actual electricity load and distributed PV generation directly impacts the accuracy of subsequent load forecasting. Existing source-load decoupling technologies include three main approaches: Hardware-based source-load decoupling: By deploying hardware such as bidirectional smart meters at the user end and dedicated data collectors for distributed photovoltaics, the output of distributed photovoltaics and the user's electricity load can be directly measured to achieve physical decoupling of source and load. This solution can directly obtain detailed data, but the hardware deployment cost is high and the cycle is long. It is difficult to implement in old power grids and low-voltage distribution networks, and it cannot achieve full network coverage measurement. Decoupling of source and load by conversion estimation: When there is no underlying subdivision measurement, the total output of distributed photovoltaic power in the whole grid is converted by combining centralized photovoltaic power output data with empirical coefficients, and then the grid-supplied net load containing disturbances is corrected. This scheme does not require additional hardware, but the conversion model does not consider the dispersion, randomness and regional irradiance differences of distributed photovoltaic power, and has inherent biases, resulting in distortion of distributed photovoltaic power output data. The load benchmark after decoupling still has significant errors. Data-driven blind source separation and decoupling: Using algorithms such as Independent Component Analysis (ICA) and Non-negative Matrix Factorization (NMF), photovoltaic output and pure electricity load are decoupled from the net grid-supplied load. This scheme requires no hardware or conversion, but it is an unconstrained data-driven model and does not incorporate prior knowledge of power system physics and meteorology. In scenarios with strong source-load correlation (such as the superposition of peak photovoltaic power generation at noon and peak air conditioning load), problems such as the aliasing of photovoltaic output and electricity load characteristics and increased decoupling error may easily occur.

[0003] Existing grid-connected net load forecasting technologies generally treat source-load decoupling as a simple preliminary step, and most schemes only output point forecast results, lacking probability quantification of load fluctuation uncertainties, making it difficult to support risk decision-making for grid dispatch. At the same time, traditional blind source separation algorithms only perform superficial decomposition of the original load sequence, failing to capture the deep characteristic correlation between photovoltaic output and electricity load, further limiting the decoupling accuracy. Summary of the Invention

[0004] The purpose of this invention is to address the problems existing in the prior art, such as heavy reliance on hardware measurement, distortion of distributed photovoltaic output conversion data, low decoupling accuracy due to the lack of physical prior constraints in traditional blind source separation, inability to capture deep source-load characteristics without introducing latent variable modeling, distortion of the overall grid load benchmark drift prediction, and lack of probability quantification capability. Combining the physical characteristics of the power system and meteorological priors, this invention provides a method and system for predicting grid net load based on latent variable decoupling and virtual metering. This method does not rely on low-level subdivided physical measurements, achieves accurate virtual metering of distributed photovoltaics based on latent variable modeling, accurate decoupling of source and load latent variables, reconstruction of the grid net load benchmark, and probabilistic prediction of source-load dynamic coupling. It solves the practical problems of increased random fluctuations and difficulty in high-precision prediction of the grid net load of provincial / municipal power grids under the background of high penetration of distributed photovoltaic power generation.

[0005] The technical solution of this invention is: A method for predicting net load of grid supply based on latent variable decoupling and virtual metering includes the following steps: S1: Data acquisition alignment and multi-scale physical feature extraction: Historical net grid load data, numerical weather forecast data and meteorological measured feature data of the target area are collected. Outlier cleaning and normalization are performed on the historical net grid load data. Variational mode decomposition is performed on the cleaned net grid load sequence to obtain multiple mode components with different center frequencies, which are divided into high-frequency fluctuation mode, mid-frequency periodic mode and low-frequency trend mode. S2: Latent variable-driven source-load blind separation. Based on the high-frequency fluctuation mode obtained in step S1 and the collected meteorological measured characteristic data, a self-encoding blind separation network is constructed to decouple the grid-supplied net load into latent variables of electricity load and latent variables of distributed photovoltaic power output. The electricity load component and the latent distributed photovoltaic power output are reconstructed by the decoder to obtain virtual metering results that do not depend on the underlying photovoltaic metering device. S3: Reconstruct the historical grid-wide electricity load benchmark sequence, use the implicit distributed photovoltaic power output obtained in step S2 to compensate the historical grid net load, obtain the grid-wide electricity load benchmark sequence after removing the influence of distributed photovoltaic self-consumption, and perform physical boundary verification. S4: Construct a dual-flow collaborative prediction network. The first path takes the benchmark sequence of the entire network's electricity load obtained in step S3 as input and predicts the benchmark value of the entire network's electricity load for the next day through a temporal convolutional network. The second path takes the meteorological characteristics of numerical weather forecast as input and predicts the implicit distributed photovoltaic output for the next day. S5: Cross-attention fusion and probability distribution correction. The latent features output by the two prediction networks in step S4 are fused through the cross-attention mechanism. The latent features of the whole grid electricity load benchmark prediction flow are used as the query vector, and the latent features of the photovoltaic disturbance prediction flow are used as the key and value vector. The reduction weight of photovoltaic disturbance on the whole grid electricity load benchmark is calculated to obtain the day-ahead net grid supply load prediction value. The multi-confidence prediction interval is output through quantile regression.

[0006] Specifically, the high-frequency fluctuation mode is: the extracted minute-level load fluctuation, which is physically mapped to the instantaneous output characteristics of distributed photovoltaic power generation affected by cloud cover and sudden changes in sunlight. The mid-frequency cyclic mode is: extracting hourly load fluctuations and physically mapping them to the diurnal operating pattern characteristics of conventional loads such as air conditioning and lighting; The low-frequency trend mode is: extracting the daily / weekly load baseline and physically mapping it to the basic electricity consumption trend characteristics brought about by socio-economic development and industrial production.

[0007] Specifically, the autoencoder blind separation network in step S2 includes a shared encoder, an electricity load mapping head, a photovoltaic mapping head, an electricity load decoder, and a photovoltaic decoder. The shared encoder maps the input features to the latent space and obtains the electricity load latent variables and distributed photovoltaic output latent variables through two independent linear mapping heads. The electricity load decoder reconstructs the electricity load components based on the electricity load latent variables and the low- and medium-frequency modes. The photovoltaic decoder reconstructs the implicit distributed photovoltaic output based on the photovoltaic latent variables and meteorological features, and the output of the photovoltaic decoder uses a non-negative constraint activation function.

[0008] Specifically, the total training loss of the autoencoder blind separation network in step S2 is composed of a weighted sum of the following sub-losses. This total training loss is used to jointly optimize the parameters of the shared encoder, the load mapper, the photovoltaic mapper, the load decoder, and the photovoltaic decoder. The reconstructed consistency loss is used to constrain the sum of the decoded electricity load component and the implicit photovoltaic output to approximate the original grid-supplied net load. Zero output loss at night / on cloudy / rainy days, used to force photovoltaic latent variables and photovoltaic decoding output to approach 0 for nighttime or low-irradiance samples; Losses related to sunny periods are used to constrain the positive correlation between distributed photovoltaic output and irradiance, and the inverse relationship with high-frequency disturbances in grid-connected net load. Latent variable decoupling loss is used to apply covariance orthogonality constraints to the latent variables of electricity load and photovoltaics. Smoothing loss is used to suppress spikes that do not conform to the continuity of photovoltaic power output.

[0009] Specifically, the physical boundary verification in step S3 includes: the benchmark for the total power load of the entire network during the nighttime period should be consistent with the original net load of the network supply; the benchmark for the total power load of the entire network during the daytime period should not be less than the original net load of the network supply; the ramp rate of adjacent sampling points should not exceed the boundary of the maximum ramp rate quantile of the same season in the historical data of the target area; if individual points exceed the boundary, they should be truncated.

[0010] Specifically, the temporal convolutional network mentioned in step S4 is an extended causal temporal convolutional network, with the expansion rate of each layer increasing in increments of 1, 2, 4, and 8. Each convolutional layer only accesses the current and historical time points. The output of the photovoltaic perturbation prediction stream is set with a non-negative constraint and multiplied by a solar irradiance gating term. When the irradiance is lower than the threshold, the output is forced to be 0.

[0011] Specifically, in step S5, the quantile regression sets up multiple quantile output heads in parallel, which output multiple sets of prediction curves including the median, 80% confidence interval and 90% confidence interval respectively; during the inference stage, monotonic correction is performed on each quantile output to ensure that the predicted value of the lower quantile is not greater than the predicted value of the higher quantile.

[0012] This invention also provides a short-term load forecasting system based on latent variable decoupling and virtual metering, comprising: The data acquisition and preprocessing module is used to perform the data acquisition alignment and multi-scale physical feature extraction steps as described in claim 1; A source-load blind separation module is used to perform the latent variable-driven source-load blind separation steps as described in claim 1; The whole network power load reconfiguration module is used to execute the steps of reconfiguring the historical whole network power load benchmark sequence as described in claim 1; A dual-stream collaborative prediction module is used to perform the construction of a dual-stream collaborative prediction network as described in claim 1; The cross-attention fusion and interval prediction module is used to perform the cross-attention fusion and probability distribution correction steps as described in claim 1.

[0013] Specifically, the data acquisition and preprocessing module includes a variational mode decomposition unit. The variational mode decomposition unit obtains K modes with different center frequencies and their center frequencies. The optimal number of modes is searched within the range of K=3 to 8, so that the reconstruction error is less than a preset threshold and the interval between adjacent center frequencies meets the minimum separation constraint. The modes are divided into high-frequency fluctuation modes, mid-frequency periodic modes and low-frequency trend modes according to the center frequency from high to low.

[0014] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the AC / DC hybrid microgrid optimization scheduling method based on multi-objective deep reinforcement learning as described above.

[0015] The beneficial effects of this invention are as follows: Virtual metering replaces physical sensing, enabling low-cost source-load decoupling and precise photovoltaic sensing. Traditional grid source-load decoupling heavily relies on hardware deployment such as bidirectional smart meters, making the transformation of existing grids difficult. The method provided in this invention utilizes "probe-based" virtual metering technology, using real-time measurable net grid load as a sensing probe. By leveraging its high-frequency fluctuation characteristics, it inversely extrapolates the implicit output of distributed photovoltaic (PV) power. This eliminates reliance on underlying hardware and solves the data distortion problem associated with centralized calculations of PV output. This suggests that in the digital transformation of the power grid, the operational side can adopt a "light asset, heavy algorithm" technical approach, combining macroscopic measurement data with algorithmic mechanisms to achieve accurate perception of microscopic distributed resources, significantly reducing the cost of upgrading grid sensing capabilities. Breaking through the limitations of traditional heavy-asset deployment of bidirectional smart meters, this innovative approach uses the real-time measurable "net grid load" as a sensing probe. Utilizing its high-frequency fluctuation characteristics, it reverse-detects and "virtually meters" the implicit electricity generated and consumed by distributed photovoltaic power in the entire grid load. This eliminates the need to convert the total distributed output through centralized photovoltaic power generation, fundamentally solving the problem of output data distortion caused by conversion, and achieving low-cost, high-precision source-load decoupling.

[0016] (2) Physical prior fusion AI algorithm breaks through the decoupling accuracy bottleneck of pure data-driven approach. Traditional blind source separation algorithms, lacking domain-specific prior constraints, are prone to feature confusion in scenarios with strong source-load correlation. This invention presents a physical prior-driven latent variable blind separation architecture, overcoming the decoupling accuracy bottleneck in scenarios with strong source-load correlation (solving the problems of traditional blind source separation lacking physical priors, not incorporating latent variable modeling, and low decoupling accuracy in scenarios with strong source-load correlation). It transforms the power system's physical and meteorological priors—such as "large photovoltaic disturbances on sunny days and no photovoltaic disturbances on rainy nights"—into mandatory constraints for the AI ​​network. A contrastive loss function for sunny and rainy weather is designed to guide the model in accurately separating deep features of electricity load and photovoltaic output features in the latent space. This suggests that in technology development, we should break away from purely data-driven algorithm design thinking and integrate the physical mechanisms of the power system as algorithmic constraints into the model. This improves the adaptability and accuracy of AI algorithms in power scenarios, significantly enhances decoupling accuracy in scenarios with strong source-load correlation, and achieves accurate source-load decoupling without hardware measurement.

[0017] (3) A new prediction paradigm based on cross-attention, namely "decoupling-reconstruction-dual-stream prediction-dynamic coupling," solves the problem of load benchmark drift. Traditional black-box prediction models directly model "polluting loads," which cannot avoid load benchmark drift caused by photovoltaic disturbances. This invention proposes a new prediction paradigm of "first decoupling source loads, reconstructing the net load benchmark of grid supply, then independently predicting source load characteristics of dual-stream systems, and finally dynamically coupling the interaction between source and loads," which fundamentally eliminates the disturbance of distributed photovoltaics on the load benchmark. It abandons the traditional black-box model of directly predicting "polluting loads." This suggests that in the modeling of complex power systems, the principle of "physical mechanism first, data modeling follow" should be followed. By restoring the true physical state of the system, a healthy foundation is laid for data-driven modeling, avoiding model bias caused by "polluting data."

[0018] (4) Probabilistic prediction of source-load dynamic coupling to support risk-driven power grid dispatching decisions Single-point forecasting is no longer sufficient to meet the needs of electricity spot trading and grid reserve capacity planning; the operational side requires quantifying the uncertainty of load fluctuations. This patent captures the nonlinear dynamic coupling relationship between source and load through cross-attention and combines it with quantile regression to output load forecast intervals with different confidence levels, thus transforming load forecasting from an "arithmetic problem" to a "risk assessment problem." This suggests that load forecasting technology should proactively adapt to the operational needs of grid dispatching, quantifying uncertainty with probability to provide data support for dispatchers in making scientific trade-offs between photovoltaic consumption and system security.

[0019] (5) A universal architecture driven by macro data, adapted to the digital upgrade of existing power grids. The method provided by this invention only requires macroscopic grid net load data and publicly available meteorological data to operate, without the need for underlying grid physical topology parameters, user-end segmented measurement data, or increased grid communication and storage pressure. This provides a strong paradigm for business application: by deploying high-order algorithm models at the dispatch center, the load perception and forecasting capabilities of the existing grid can be upgraded, maximizing the value of existing grid data assets and accelerating the implementation of grid digital transformation. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the method provided in Example 1. Detailed Implementation

[0021] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0022] Example 1

[0023] This embodiment provides a short-term load forecasting method based on latent variable decoupling and virtual metering. The target area can be a provincial power grid, a municipal power grid, or a region encompassing several substations / feeders; the target output is a 96-point load forecast sequence (including total grid load and net grid supply load) arranged at 15-minute intervals from 0:00 to 24:00 the next day, simultaneously providing multi-confidence forecast intervals. For ease of explanation, the sampling interval is assumed to be Δt = 15 min, the number of sampling points per day is T = 96, and the historical window length is D. L =30 days, the length of the daytime weather window D W =7 days.

[0024] In this embodiment, the directly measurable data includes the historical net load sequence of the target area. Numerical Weather Prediction (NWP) Data Calendar features and necessary regional installation ledgers . This represents the net grid load visible on the dispatch side at the t-th sampling point on day d; due to the self-consumption of distributed photovoltaic power, The numerical value is not equal to the actual electricity load on the user side, but rather the remaining electricity load after offsetting by distributed photovoltaic power generation. The core of this invention is to deduce the possible distributed photovoltaic power generation through analysis of electrical and meteorological data without adding new physical electricity meters. And on this basis, the actual electricity load is reconstructed, that is .

[0025] To avoid confusion in terminology, the following conventions are used in this embodiment: This represents the net grid supply load visible to the dispatching side, which is the net grid supply load after being offset by distributed photovoltaic self-consumption. This represents the distributed photovoltaic power generation obtained through virtual metering inversion, which is also the implicit deduction from the actual electricity load. Indicates will Add back The resulting nationwide power load benchmark is used to characterize a load level that more closely reflects users' actual power demand. Therefore, the historical reconstruction phase adopts... The output stage is currently predicted to use The directions of addition and subtraction are the same for both.

[0026] The three sequences mentioned above are all arranged along a unified time axis. By default, there are 96 sampling points per day (T=96), where d represents the date index and t represents the index of the sampling point within 15 minutes of the day. When the sampling interval is changed to 5 minutes, 30 minutes, or hourly, only T and the input / output dimensions need to be adjusted synchronously, without changing the technical logic of source-load decoupling, virtual metering, and dual-stream fusion.

[0027] Historical net load available from the dispatching side d represents the date. The benchmark for total grid power load after compensating for implicit photovoltaic output Same dimension Implicit output of distributed photovoltaic power obtained by decoupling from implicit variables The value should be close to 0 at night and non-negative during the day. Meteorological characteristics such as irradiance, cloud cover, temperature, humidity, and wind speed. Interpolation from hourly levels to 96 points is possible. Hidden variables of electricity load and distributed photovoltaic output encoder latent space vector Queries, keys, and values ​​in cross attention Obtained by linear mapping of two-way prediction features

[0028] The specific steps of the method provided in this embodiment are as follows: S1: Data Acquisition Alignment and Multi-Scale Physical Feature Extraction. Historical net grid load data, numerical weather prediction data, and meteorological observation data of the target area are collected. Outlier cleaning and normalization are performed on the historical net grid load data. Variational mode decomposition is then applied to the cleaned net grid load sequence to obtain multiple modal components with different center frequencies, categorized as high-frequency fluctuation modes, mid-frequency periodic modes, and low-frequency trend modes. In the model training and inference process, this embodiment follows a closed-loop procedure: First, the source-load blind separation network in S2 is trained using historical net grid load and historical meteorological data to obtain historical daily data. and Then with As a flow label for predicting the electricity load of the entire network, As a label for photovoltaic disturbance prediction flow, the dual-flow prediction network in S4 is trained; then, historical real grid supply net load is used. As supervisory labels, the cross-attention fusion layer and quantile prediction head in S5 are trained; during online inference, only the historical data before the target date is input. Based on the current NWP and calendar features, the following day's output can be generated. , , And the prediction range. This process is explained. The training labels are derived from the virtual metering results of S2, rather than relying on newly added underlying distributed photovoltaic metering devices. Specifically, the steps are as follows: S1.1 Data Acquisition and Time Alignment. Collect historical net network load data for the target area for at least the most recent 30 days, with a uniform load sampling interval of 15 minutes; collect numerical weather forecasts or meteorological observation data corresponding to the target area, with meteorological elements including at least one or more of the following: total horizontal irradiance, cloud cover, temperature, and humidity. If the meteorological data is hourly, interpolate it to a 15-minute granularity along the time axis and align it with the load sampling points.

[0029] S1.2 Outlier Cleaning. For missing points in the load sequence, if the consecutive missing length does not exceed 4 sampling points, linear interpolation of adjacent valid points is used; if the consecutive missing length exceeds 4 sampling points, the mean of samples from the same week, same hour, and similar temperatures is used for imputation. For peak outliers, a local window is calculated. median within and median absolute deviation When satisfied If there is no slope support in the same direction between adjacent sampling points, the point is identified as an anomaly and replaced with a local smooth value.

[0030]

[0031] in, and These are the most recent valid sampling points before and after the missing point. The above cleaning is only used to remove non-business disturbances such as communication gaps and measurement spikes, and does not erase real physical fluctuations such as photovoltaic output, temperature changes, and load ramp-up.

[0032] S1.3 Normalization and Feature Construction. Continuous variables such as load, irradiance, temperature, and humidity are standardized using training set statistics. Calendar-type variables are standardized using one-hot encoding or 0-1 encoding. The load standardization formula is as follows:

[0033] in, and These represent the mean and standard deviation of the net load for the training network, respectively. Meteorological characteristics. Before entering the network, the mean and standard deviation of the training set are also used to avoid model bias caused by directly splicing features of different dimensions.

[0034] S1.4 performs improved VMD multi-scale decomposition. This is applied to the cleaned net network load sequence. Variational mode decomposition is performed to obtain K modes with different center frequencies. and center frequency The optimization objective of VMD is to reconstruct the original load sequence by superimposing the modal bandwidths, while minimizing the sum of the bandwidths of all modalities.

[0035] In practice, K is not fixed; it can be searched within the range of K=3 to 8 first to minimize the reconstruction error. The frequency must be less than a preset threshold (e.g., 1%), and the interval between adjacent center frequencies must satisfy the minimum separation constraint. After decomposition, the frequencies are divided into high-frequency fluctuation modes according to their center frequencies from high to low. Mid-frequency periodic mode and low-frequency trend modes :

[0036] Among them, high-frequency fluctuation mode The main characterization features include transient disturbances in distributed photovoltaic systems caused by cloud cover and abrupt changes in irradiance; mid-frequency periodic modes. Primarily characterizes intraday rhythms of air conditioning, lighting, and industrial production; low-frequency trend modes. It primarily characterizes the cyclical, seasonal, and baseline electricity consumption trends. This mapping serves only as a physical prior for subsequent latent variable decoupling; it does not simply equate high-frequency modes with photovoltaic output, but rather inputs them into the encoder along with meteorological priors for constraint learning.

[0037] S2: Latent variable-driven source-load blind separation. Based on the high-frequency fluctuation modes obtained in step S1 and the collected meteorological measurement characteristic data, a self-encoding blind separation network is constructed to decouple the grid-supplied net load into implicit variables of electricity load and distributed photovoltaic output. The electricity load component and implicit distributed photovoltaic output are reconstructed by the decoder, respectively, to obtain virtual metering results that do not depend on the underlying photovoltaic metering device. Specifically, the following steps are included: S2.1 encoder input construction. High-frequency fluctuation modes. Local slope rate Irradiance Cloud cover ,temperature ,humidity Sunrise and sunset signs Features are concatenated at different time points to form a source-load decoupled input vector. .in Determined by the solar altitude angle or irradiance threshold, it is 0 during nighttime and cloudy / rainy periods with low irradiance, and 1 during sunny periods with sunshine.

[0038]

[0039] S2.2 Autoencoder Blind Separation Network. Constructing the encoder. Power load decoder and photovoltaic power decoder The encoder will Mapped to the latent space, and the latent variables of electricity load are obtained through two independent linear mapping heads. Hidden variables of distributed photovoltaic power output :

[0040] According to the power load decoder Mid- and low-frequency modes 、 Restore the implicit component of electrical load; photovoltaic decoder according to The implicit distributed photovoltaic (PV) output is reconstructed based on meteorological characteristics. To ensure correct physical direction, the PV output decoding end uses a non-negative constraint activation function, such as Softplus or ReLU, to ensure the output satisfies... .

[0041]

[0042] In a specific and implementable network architecture, the encoder A three-layer multilayer perceptron is used, with the input dimension being... The number of features, the hidden layer dimension can be set to 64 or 32, and the output latent space dimension can be set to 16; power load mapping head. and photovoltaic mapping head Output 8-dimensional hidden variables respectively and Using an electrical load decoder Using a two-layer, multi-layer perceptron, , , After splicing, it is mapped to a value with load dimensions. The photovoltaic decoder (DPV) uses a two-layer, multi-layer sensor to... With meteorological characteristics After concatenation, the mapping is as follows The output layer uses Softplus or ReLU to ensure non-negativity. The number of layers and dimensions can be adjusted according to the sample size, but the structure of "shared encoding, dual-head latent variables, dual decoders, and physical constraint loss" should be retained.

[0043] S2.3 Reconstructing Consistency Constraints. Since the net grid load observed by the dispatching side equals the actual user electricity load minus the implicit distributed photovoltaic output, the decoded electricity load component and the implicit photovoltaic output are required to satisfy the following power balance relationship during training:

[0044] This constraint ensures that the model does not arbitrarily split the two latent variables, but rather that the synthesis of the two latent variables can return to the actual observed net load curve of the grid supply.

[0045] S2.4 Zero output constraint at night / dark / rainy weather. For nighttime or low-irradiance sample sets. Force the photovoltaic hidden variable and the photovoltaic decoding output to approach 0: .

[0046] S2.5 High-frequency correlation constraints during sunny periods. For the set of samples with sunshine. Distributed photovoltaic output should have a stable correlation with the high-frequency inverse fluctuations of irradiance and grid-supplied net load. Therefore, correlation constraints are introduced to enhance [the effectiveness of this correlation]. Capability to express photovoltaic disturbances:

[0047] Where Corr(·) represents the batch-specific Pearson correlation coefficient. Since distributed photovoltaic self-consumption reduces the net grid load, during sunny periods... It should be positively correlated with irradiance and inversely correlated with the high-frequency disturbance of the grid-connected net load dip caused by photovoltaics; the above constraints are used to prevent the model from misidentifying the midday photovoltaic reduction as a natural decrease in electricity load.

[0048] S2.6 Latent Variable Decoupling Constraints. To reduce information aliasing between the latent variables of electricity load and photovoltaic power, covariance orthogonality constraints are introduced for both:

[0049] S2.7 Total loss for source-load blind separation. The model training loss is composed of reconstruction consistency, zero-output prior, clear correlation prior, latent variable orthogonality, and output smoothness.

[0050] in, These are weighting coefficients, which can be determined using the validation set; This represents the second-order difference, used to suppress spikes that do not conform to the continuity of photovoltaic power output. After training, This refers to the implicit distributed photovoltaic virtual metering results that do not rely on the underlying dedicated photovoltaic metering device. The implicit component of the electrical load that is matched with it.

[0051] S3: Reconstruct the historical grid-wide electricity load baseline sequence, and use the implicit distributed photovoltaic output obtained in step S2 to compensate for the historical grid net load, obtaining the grid-wide electricity load baseline sequence after removing the influence of distributed photovoltaic self-consumption, and perform physical boundary verification:

[0052] The above compensation direction means: net load supplied by the grid. This is the visible load after the actual user load is offset by distributed photovoltaic self-consumption; therefore, the virtual metering obtained... After adding it back, we get a result that more closely reflects the user's actual electricity needs. .

[0053] To avoid overcompensation, physical boundary checks are also performed after reconstruction: at night Should be with Basically the same; daytime It should be no less than The slope rate of adjacent sampling points should not exceed the boundary of the maximum slope quantile in the same season in the historical data of the target area. The verification rules are as follows:

[0054] in, This represents the 99.5th percentile of the absolute value of the historical climb rate. If any individual point exceeds the boundary, it is truncated according to the boundary, and the truncation amount is fed back into the smoothing loss weight adjustment of S2. The result obtained through this step... The sequence serves as a training label and input base for subsequent nationwide electricity load forecasting, thereby reducing load baseline drift caused by changes in distributed photovoltaic penetration rate at the source.

[0055] S4: Construct a dual-flow collaborative prediction network. The first path uses the baseline sequence of the entire network's electricity load obtained in step S3 as input, and predicts the baseline value of the entire network's electricity load for the next day through a temporal convolutional network. The second path uses the meteorological characteristics of numerical weather forecasts as input, and predicts the implicit distributed photovoltaic output for the next day. Specifically, it includes the following steps: S4.1 Sample Window Construction. For each target prediction day D+1, two inputs are constructed: the first is the reconstructed baseline sequence of the entire network's electricity load over the past 30 days. The first path is used to learn the periodicity and trends of real-world electricity demand; the second path contains numerical weather forecast meteorological characteristics for the past 7 days and the target day. This is used to learn about distributed photovoltaic power output disturbances. Date features include weekday, month, public holidays, work schedule adjustment type, and season identifier.

[0056] S4.2 Overall Network Electricity Load Baseline Forecast Flow. This flow employs an extended causal temporal convolutional network (TCN). Each convolutional layer only accesses current and historical time points, avoiding the use of future load information; the expansion rate increases in increments of 1, 2, 4, and 8, allowing the model to cover cross-day and cross-week dependencies with a fewer number of layers. The output of the l-th layer is:

[0057] Where k is the kernel width, Let l be the expansion rate of the l-th layer, and the residual term be... Used to maintain gradient stability. The total grid electricity load forecast outputs the baseline forecast value for the total grid electricity load at 96 points the following day. .

[0058] S4.3 Distributed Photovoltaic Disturbance Prediction Stream. This stream takes day-ahead irradiance, cloud cover, temperature, humidity, and sunshine duration as inputs and outputs an implicit distributed photovoltaic power output prediction for the following day. To maintain consistency with the S2 virtual econometric results, the training labels were obtained from the historical daily inversion of S2. Instead of a centralized photovoltaic conversion value, this predicted flow can be achieved using a lightweight TCN or a multilayer sensor, with its output also subject to non-negativity constraints and multiplied by a solar irradiance gating term.

[0059] in, The output is forced to be 0 at night to avoid generating false photovoltaic output at night; The output intensity is determined by meteorological conditions.

[0060] S4.4 Dual-stream training loss. To ensure that the two forecasts respectively align with the grid-wide electricity load benchmark and photovoltaic disturbances, a weighted loss method is used for training:

[0061] Among them, Huber loss is used to reduce the interference of abnormal load days on the overall grid power load forecast, MSE is used to constrain the accuracy of photovoltaic disturbance forecast, and ramp-up loss term... It is used to enhance the fitting ability of morning and evening peak climbing and midday concave morphology.

[0062] S5: Cross-attention fusion and probability distribution correction. The latent features output from the two prediction networks in step S4 are fused using a cross-attention mechanism. The latent features of the total grid electricity load baseline prediction flow are used as the query vector, and the latent features of the photovoltaic disturbance prediction flow are used as the key-value vector. The reduction weight of the photovoltaic disturbance on the total grid electricity load baseline is calculated to obtain the day-ahead net grid load prediction value. A multi-confidence prediction interval is then output through quantile regression. Specifically, the steps are as follows: S5.1 Cross-Attention Fusion. Latent characteristics of the predicted flow output of the entire network's electricity load. The latent features of the photovoltaic perturbation prediction stream output as the query vector Q. Using K as the key and V as the value, the reduction weight of photovoltaic disturbances on the benchmark electricity load of the entire grid at different time periods is calculated through cross-attention:

[0063]

[0064] Where A represents the dynamic coupling weight between time periods; the minus sign reflects the reduction effect of distributed photovoltaic power on the overall grid electricity load benchmark in the grid net load forecast. Based on Preliminary day-ahead net load forecast is obtained from the output layer:

[0065] Wherein, γ(t) is determined by cross-attention weights and solar irradiance gating, and its value ranges from [0,1] or is calibrated according to the self-consumption ratio of distributed photovoltaic power in the target area. This expression maintains the technical logic of this patent: "first reconstructing the benchmark of the entire grid electricity load, and then subtracting the predicted photovoltaic disturbances to form the day-ahead grid supply / net grid supply load prediction".

[0066] It should be noted that in cross-attention... This is used to learn the dynamic coupling relationship between electricity load forecasting characteristics and photovoltaic disturbance forecasting characteristics, and to generate or correct γ(t) accordingly. Its function is to determine the effective reduction ratio of grid-connected net load due to photovoltaic disturbances at different time periods; the deduction in physical dimensions only applies to... This process is performed only once to avoid repeatedly subtracting the same photovoltaic perturbation in the feature layer and the output layer.

[0067] when When defined as the implicit deduction of net grid load due to self-consumption, γ(t) can take a correction factor close to 1 and be finely adjusted by attention weights; when When the definition is expanded to the total output of distributed photovoltaic power in the target area, This indicates the proportion of self-generated and self-consumed power that actually affects the net load of the grid, or the equivalent reduction proportion. This invention adopts the former definition, namely... This is an implicit deduction. Its main function is to perform time-phase coupling correction.

[0068] S5.2 Quantile Regression Interval Correction. To output probability prediction results, multiple quantile output heads are set up in parallel outside the point prediction output layer, each outputting its own probability prediction result. Corresponding prediction curve The quantile loss is:

[0069] in, As the predicted median value to Form a 90% confidence interval. to An 80% confidence interval is formed. To avoid quantile crossover, monotonic correction is performed on the output of each quantile during the inference phase, so that... .

[0070] S5.3 Final Output. The final output includes: the expected net load forecast curve for 96 points of the power grid for the next day. 1. Baseline forecast curve of total grid electricity load Implicit Distributed Photovoltaic Disturbance Prediction Curve And confidence intervals such as 80% and 90%. The scheduling side can use... You can also use it to submit day-ahead plans and spot transaction declarations. Identify changes in actual electricity demand and use To help determine the impact of distributed photovoltaic power on midday lows and evening uphill climbs.

[0071] S1 Historical network provides net load, NWP, and calendar features. Anomaly cleaning, normalization, VMD multiscale decomposition High-frequency / mid-frequency / low-frequency modes and normalized characteristics S2 High-frequency modes, meteorological priors, and net grid load Self-encoding latent space decoupling, zero-output / correlation / orthogonality constraints 、 S3 、 Compensate and perform physical verification. Historical power load benchmark S4 Historical Window, NWP (Nearby) Parallel training of electrical load TCN predicted current and photovoltaic disturbance predicted current. 、 S5 Two-way prediction of latent features and output Cross-attention deduction, quantile regression interval correction The following day's forecast values ​​for the net load of the power grid at 96:00 AM, the baseline forecast value for the total power load of the entire grid, the forecast value for the implicit deduction of photovoltaic power, and multiple confidence intervals. The method provided in this embodiment has the following advantages: 1. It compensates for grid-connected net load pollution by replacing large-scale underlying hardware measurement with implicit distributed photovoltaic virtual metering without changing the existing scheduling data acquisition system; 2. It provides a clear physical basis for source-load decoupling through zero output at night, clear weather correlation, orthogonality of latent variables, and reconstruction consistency constraints, avoiding feature aliasing of blind data separation; 3. By first reconstructing the grid-wide electricity load benchmark and then performing dual-flow prediction, the model training object is transformed from the pollution curve affected by self-generation and self-consumption to a more stable real electricity demand curve; 4. By simultaneously outputting point prediction and interval prediction through cross-attention and quantile regression, the prediction results can support day-ahead scheduling, reserve capacity arrangement, and spot trading risk assessment.

[0072] Example 2 This embodiment provides a whole-network pure network net load forecasting system based on latent variable decoupling and virtual metering, including a data acquisition alignment and multi-scale physical feature extraction module, a latent variable-driven source-load blind separation module, a historical benchmark sequence reconstruction module, a dual-stream collaborative day-ahead forecasting module, and a cross-attention fusion and probability distribution correction module.

[0073] The data acquisition alignment and multi-scale physical feature extraction module receives historical net network load, NWP meteorological data, calendar features, and regional installed capacity ledgers from SCADA or the dispatch master station. This module performs timestamp alignment, outlier cleaning, normalization, and VMD multi-scale decomposition, and outputs a standardized feature matrix. , , And the corresponding physical tags.

[0074] The latent variable-driven source-load blind separation module is used to execute step S2 of Example 1. This module internally includes an encoder, an electrical load decoder, a photovoltaic output decoder, and a physical constraint loss calculation unit; the encoder output... and Decoder output and The loss calculation unit updates the network parameters based on reconstruction consistency, zero output at night, clear weather correlation, orthogonality of latent variables, and smoothing constraints.

[0075] The historical baseline sequence reconstruction module is used to perform step S3 of Example 1. This module will... Add back historical net load ,generate And perform physical verification based on rules such as non-negative compensation, nighttime consistency, and climbing boundaries; after verification... Enter the model training sample library.

[0076] The dual-stream coordinated day-ahead forecasting module is used to perform step S4 of Example 1. This module includes an electricity load baseline forecasting stream and a distributed photovoltaic disturbance forecasting stream; the former uses... The history window is for input and output. The latter takes the current NWP and Rizhao gating as inputs and outputs... The two prediction results are strictly aligned along a 96-point timeline.

[0077] The cross-attention fusion and probability distribution correction module is used to execute step S5 of Example 1. This module calculates the cross-attention weights based on the implicit features of electricity load prediction and photovoltaic disturbance prediction, forming... It outputs different confidence intervals through the quantile prediction head; the output interface includes at least the prediction date, the predicted value of the net load of the 96-point grid supply, the upper and lower boundaries of the 96-point interval, the benchmark curve of the total grid electricity load, the implicit photovoltaic deduction curve, and the model version number.

[0078] The system can be deployed at provincial or municipal dispatch centers and interacts with existing load forecasting platforms through file interfaces, database interfaces, or API interfaces. During system operation, no new user-side distributed photovoltaic metering devices are required; the system primarily relies on existing grid-supplied net load, weather forecasts, and calendar data to complete source-load decoupling, virtual metering, and day-ahead forecasting.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for predicting net load of power grid supply based on latent variable decoupling and virtual metering, characterized in that, Includes the following steps: S1: Data acquisition alignment and multi-scale physical feature extraction: Historical net grid load data, numerical weather forecast data and meteorological measured feature data of the target area are collected. Outlier cleaning and normalization are performed on the historical net grid load data. Variational mode decomposition is performed on the cleaned net grid load sequence to obtain multiple mode components with different center frequencies, which are divided into high-frequency fluctuation mode, mid-frequency periodic mode and low-frequency trend mode. S2: Latent variable-driven source-load blind separation. Based on the high-frequency fluctuation mode obtained in step S1 and the collected meteorological measured characteristic data, a self-encoding blind separation network is constructed to decouple the grid-supplied net load into latent variables of electricity load and latent variables of distributed photovoltaic power output. The electricity load component and the latent distributed photovoltaic power output are reconstructed by the decoder to obtain virtual metering results that do not depend on the underlying photovoltaic metering device. S3: Reconstruct the historical grid-wide electricity load benchmark sequence, use the implicit distributed photovoltaic power output obtained in step S2 to compensate the historical grid net load, obtain the grid-wide electricity load benchmark sequence after removing the influence of distributed photovoltaic self-consumption, and perform physical boundary verification. S4: Construct a dual-flow collaborative prediction network. The first path takes the benchmark sequence of the entire network's electricity load obtained in step S3 as input and predicts the benchmark value of the entire network's electricity load for the next day through a temporal convolutional network. The second path takes the meteorological characteristics of numerical weather forecast as input and predicts the implicit distributed photovoltaic output for the next day. S5: Cross-attention fusion and probability distribution correction. The latent features output by the two prediction networks in step S4 are fused through the cross-attention mechanism. The latent features of the whole grid electricity load benchmark prediction flow are used as the query vector, and the latent features of the photovoltaic disturbance prediction flow are used as the key and value vector. The reduction weight of photovoltaic disturbance on the whole grid electricity load benchmark is calculated to obtain the day-ahead net grid supply load prediction value. The multi-confidence prediction interval is output through quantile regression.

2. The method for predicting net load of the entire power grid according to claim 1, characterized in that, The high-frequency fluctuation mode is: the extracted minute-level load fluctuation, which is physically mapped to the instantaneous output characteristics of distributed photovoltaic power generation affected by cloud cover and sudden changes in sunlight. The mid-frequency cyclic mode is: extracting hourly load fluctuations and physically mapping them to the diurnal operating pattern characteristics of conventional loads such as air conditioning and lighting; The low-frequency trend mode is: extracting the daily / weekly load baseline and physically mapping it to the basic electricity consumption trend characteristics brought about by socio-economic development and industrial production.

3. The method for predicting net load of the entire power grid according to claim 1, characterized in that, The autoencoder blind separation network in step S2 includes a shared encoder, an electricity load mapping head, a photovoltaic mapping head, an electricity load decoder, and a photovoltaic decoder. The shared encoder maps the input features to the latent space and obtains the electricity load latent variables and distributed photovoltaic output latent variables through two independent linear mapping heads. The electricity load decoder reconstructs the electricity load components based on the electricity load latent variables and the low- and medium-frequency modes. The photovoltaic decoder reconstructs the implicit distributed photovoltaic output based on the photovoltaic latent variables and meteorological features, and the output of the photovoltaic decoder uses a non-negative constraint activation function.

4. The method for predicting net load of the entire power grid according to claim 3, characterized in that, The total training loss of the autoencoder blind separation network in step S2 is composed of a weighted sum of the following sub-losses. This total training loss is used to jointly optimize the parameters of the shared encoder, the load mapper, the photovoltaic mapper, the load decoder, and the photovoltaic decoder. The reconstructed consistency loss is used to constrain the sum of the decoded electricity load component and the implicit photovoltaic output to approximate the original grid-supplied net load. Zero output loss at night / on cloudy / rainy days, used to force photovoltaic latent variables and photovoltaic decoding output to approach 0 for nighttime or low-irradiance samples; Losses related to sunny periods are used to constrain the positive correlation between distributed photovoltaic output and irradiance, and the inverse relationship with high-frequency disturbances in grid-connected net load. Latent variable decoupling loss is used to apply covariance orthogonality constraints to the latent variables of electricity load and photovoltaics. Smoothing loss is used to suppress spikes that do not conform to the continuity of photovoltaic power output.

5. The method for predicting net load of the entire power grid according to claim 1, characterized in that, The physical boundary verification in step S3 includes: the benchmark for the total power load of the network during the nighttime period should be consistent with the original net load of the network supply; the benchmark for the total power load of the network during the daytime period should not be less than the original net load of the network supply; the ramp rate of adjacent sampling points should not exceed the boundary of the maximum ramp rate quantile of the same season in the historical data of the target area; if individual points exceed the boundary, they should be truncated.

6. The method for predicting net load of the entire power grid according to claim 1, characterized in that, The temporal convolutional network mentioned in step S4 is an extended causal temporal convolutional network, with the expansion rate of each layer increasing in increments of 1, 2, 4, and 8. Each convolutional layer only accesses the current and historical time points. The output of the photovoltaic disturbance prediction stream is set with a non-negative constraint and multiplied by a solar irradiance gating term. When the irradiance is lower than the threshold, the output is forced to be 0.

7. The method for predicting net load of the entire power grid according to claim 1, characterized in that, In step S5, the quantile regression is set up in parallel with multiple quantile output heads, which output multiple sets of prediction curves including the median, 80% confidence interval and 90% confidence interval respectively; during the inference stage, monotonic correction is performed on each quantile output to ensure that the predicted value of the lower quantile is not greater than the predicted value of the higher quantile.

8. A short-term load forecasting system based on latent variable decoupling and virtual metering, characterized in that, include: The data acquisition and preprocessing module is used to perform the data acquisition alignment and multi-scale physical feature extraction steps as described in claim 1; A source-load blind separation module is used to perform the latent variable-driven source-load blind separation steps as described in claim 1; The whole network power load reconfiguration module is used to execute the steps of reconfiguring the historical whole network power load benchmark sequence as described in claim 1; A dual-stream collaborative prediction module is used to perform the construction of a dual-stream collaborative prediction network as described in claim 1; The cross-attention fusion and interval prediction module is used to perform the cross-attention fusion and probability distribution correction steps as described in claim 1.

9. The system according to claim 8, characterized in that, The data acquisition and preprocessing module includes a variational mode decomposition unit. The variational mode decomposition unit obtains K modes with different center frequencies and their center frequencies. The optimal number of modes is searched within the range of K=3 to 8 to make the reconstruction error less than a preset threshold and the interval between adjacent center frequencies meet the minimum separation constraint. The modes are divided into high-frequency fluctuation modes, mid-frequency periodic modes and low-frequency trend modes according to the center frequency from high to low.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the AC / DC hybrid microgrid optimization scheduling method based on multi-objective deep reinforcement learning as described in any one of claims 1 to 7.

Images