Power grid net load probability prediction method and system based on conditional boltzmann machine

Abstract

The present application belongs to the technical field of power system load forecasting, and particularly relates to a power grid net load probability forecasting method and system based on a conditional Boltzmann machine. The method comprises the following steps: binary coding of the normalized power system net load into a binary visible layer vector; constructing a continuous condition vector based on historical load data of the power system; changing the condition energy function in the conditional Boltzmann machine only by modulating the visible bias and the hidden bias of the continuous condition vector; training the parameters of the conditional Boltzmann machine, and obtaining the empirical prediction distribution and the continuous grading probability score of the target net load to be predicted based on the trained conditional Boltzmann machine. The present application injects the continuous condition feature into the condition energy function only by modulating the visible bias and the hidden bias, so that the order of magnitude of the continuous condition is absorbed by the bias term, and the weight only needs to adapt to the interaction between the binary layers, thereby fundamentally eliminating the order mismatch.

Description

Technical Field

[0001] This invention belongs to the field of power system load forecasting technology, and particularly relates to a method and system for predicting the net load of a power grid based on a conditional Boltzmann machine. Background Technology

[0002] In new power systems, the high proportion of renewable energy integration significantly increases the volatility of net load. Intraday dispatching (especially congestion management and reserve arrangements) requires not only point estimates of net load but also its probability distribution to support risk-based decision-making. In existing technologies, mainstream deterministic forecasting methods only output a single point value, which cannot characterize forecast uncertainty and is insufficient to support dispatching decisions that require distributed information.

[0003] Energy-based probabilistic models (such as Restricted Boltzmann Machines, RBMs) can naturally output distributions, but their visible layer of energy functions is defined in binary space. If the standardized continuous net load is directly input into the binary visible layer, it will cause a mismatch between the continuous and binary orders of magnitude. Continuous values Visible cell activation and weights are forcibly assigned to visible cells that can only take the values ​​0 or 1. The inner product must simultaneously accommodate two incompatible scaling requirements: "real-order scale of continuous input" and "binary scale of Boltzmann distribution." During training, the weights... To compensate for this scale conflict, the model continues to amplify and grow unbounded, failing to converge to an effective probability distribution. Experimental observations confirm that when continuous loads are directly input into the binary energy model for training, the weight norm continues to grow unbounded during training, and the point prediction error saturates and no longer improves.

[0004] In addition, power grid load forecasting is highly dependent on high-dimensional continuous exogenous features such as meteorology, but traditional RBMs lack a mechanism to inject continuous heterogeneous features into the energy model without destroying its binary structure and causing the aforementioned magnitude mismatch. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a method and system for predicting the net load of a power grid based on a conditional Boltzmann machine. While maintaining the binary structure of the visible layer, continuous conditional features are injected into the conditional energy function by modulating the visible bias and the hidden bias. This is achieved by two sets of independent learnable linear mappings, so that the real order of magnitude of the continuous condition is absorbed by the bias term and the weights only need to be adapted to the interaction between the binary layers, thus fundamentally eliminating the order of magnitude mismatch.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for predicting the net load probability of a power grid based on a conditional Boltzmann machine.

[0007] The method for probabilistic prediction of grid net load based on conditional Boltzmann machines includes the following steps: The standardized net load of the power system is binary-coded into a binary visible layer vector; Constructing continuous conditional vectors based on historical load data of the power system; This allows the continuous conditional vector to change the conditional energy function in the conditional Boltzmann machine only by modulating the visible and hidden biases; The parameters of the conditional Boltzmann machine are trained, and based on the trained conditional Boltzmann machine, the empirical predicted distribution and continuous graded probability score of the target net load are obtained.

[0008] The second aspect of the present invention provides a power grid net load probability prediction system based on a conditional Boltzmann machine.

[0009] A grid net load probabilistic prediction system based on a conditional Boltzmann machine includes: The sequence-preserving binary encoding module is configured to: encode the standardized net load of the power system into a binary visible layer vector; The continuous condition construction module is configured to construct a continuous condition vector based on historical load data of the power system. The bias modulation module is configured to enable the continuous condition vector to change the conditional energy function in the conditional Boltzmann machine only by modulating the visible bias and the hidden bias; The training and prediction module is configured to: train the parameters of the conditional Boltzmann machine, and based on the trained conditional Boltzmann machine, obtain the empirical prediction distribution and continuous hierarchical probability score of the target net load to be predicted.

[0010] The above one or more technical solutions have the following beneficial effects: This invention provides a method and system for predicting the net load of a power grid based on a conditional Boltzmann machine. While maintaining the binary structure of the visible layer, the continuous conditional features are injected into the conditional energy function by modulating the visible bias and the hidden bias. This is achieved by two sets of independent learnable linear mappings, so that the real order of magnitude of the continuous condition is absorbed by the bias term and the weights only need to be adapted to the interaction between the binary layers, thus fundamentally eliminating the order of magnitude mismatch.

[0011] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0013] Figure 1 This is a flowchart illustrating the steps of Embodiment 1 of the present invention.

[0014] Figure 2 This is a data flow diagram of Embodiment 1 of the present invention.

[0015] Figure 3 This is a measured bar chart of the continuous graded probability score of CRBM on typical days in three seasons (winter heavy / waist lotus / summer heavy) according to Embodiment 1 of the present invention. Detailed Implementation

[0016] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0017] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0018] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0019] Example 1 To address the issues of training divergence caused by the continuous-binary magnitude mismatch and the difficulty in utilizing high-dimensional continuous conditions, this invention proposes a probability prediction method for grid net load based on a conditional Boltzmann machine. This method encodes the per-unit net load into a binary visible layer vector using a thermometer-preserved encoding method. Continuous condition features are injected into the conditional energy function by modulating the visible and hidden biases using two independent linear mappings, respectively. This allows the real magnitude of the continuous conditions to be absorbed by the bias term, and the weights only need to adapt to the binary layer interactions, thus resolving the magnitude mismatch when the energy model handles continuous variables at the architectural level. All parameters are jointly trained end-to-end using the contrastive divergence loss of the conditional free energy difference. A complete empirical prediction distribution is generated through preheating and multiple conditional Gibbs sampling, and a continuous hierarchical probability score is calculated. The training of this invention is stable (the standard deviation of the weights is stable at about 0.23). On a typical day of the three seasons in a certain provincial power grid, the CRPS (Continuous Ranked Probability Score) is 0.0098 / 0.0092 / 0.0225. It can support risk-based dispatching decisions such as intraday congestion management and reserve arrangements in the power grid. The output empirical prediction distribution and continuous ranked probability score can be used by dispatchers or downstream decision-making modules.

[0020] As shown in Figure 1, the grid net load probabilistic prediction method based on the conditional Boltzmann machine includes the following steps: The standardized net load of the power system is binary-coded into a binary visible layer vector; Constructing continuous conditional vectors based on historical load data of the power system; This allows the continuous conditional vector to change the conditional energy function in the conditional Boltzmann machine only by modulating the visible and hidden biases; The parameters of the conditional Boltzmann machine are trained, and based on the trained conditional Boltzmann machine, the empirical predicted distribution and continuous graded probability score of the target net load are obtained.

[0021] This embodiment constructs a continuous condition vector. As conditional input, continuous conditional vector Visible bias in the modulation conditional energy function only With hidden bias It participates in modeling in a manner that does not interact with the activation state of the binary visible layer or the activation state of the hidden layer through multiplication.

[0022] Therefore, while maintaining the binary structure of the visible layer, this invention injects continuous conditional features into the conditional energy function by modulating the visible bias and the hidden bias, which is implemented by two sets of independent learnable linear mappings. This allows the real order of magnitude of the continuous condition to be absorbed by the bias term, and the weights only need to be adapted to the interaction between the binary layers, thus fundamentally eliminating the order of magnitude mismatch.

[0023] The proposed solution in this embodiment will now be explained in detail.

[0024] The grid net load probabilistic prediction method based on conditional Boltzmann machines provided in this embodiment consists of four mutually cooperating steps: order-preserving binary encoding, double-bias conditional injection, conditional contrastive divergence training, and conditional sampling probability prediction. These four steps are not simply pieced together. Order-preserving binary encoding maintains the binary structure of the energy model. Double bias injection introduces high-dimensional continuous conditions and resolves magnitude conflicts without destroying the binary structure. Conditional contrastive divergence training trains all parameters end-to-end. Conditional sampling produces a complete empirical distribution. The absence of any link (such as changing continuous features to directly enter the visible layer or changing order-preserving encoding to non-order-preserving one-hot encoding) will cause the method to degenerate or fail.

[0025] I. Sequence-preserving binary encoding.

[0026] Standardized net load The thermometer encoding is discretized into a binary visible layer vector:

[0027] in, The number of units that are set to 1; This indicates the rounding operation; This represents the net load of the standardized power system; Indicates the adjustable quantization level; This represents the binary visible layer vector of the thermometer's encoded output, which is the ordered binary vector mapped after quantization, with the first k cells set to 1 and the rest set to 0.

[0028] It can be seen that the layer vector exhibits a monotonically non-increasing or monotonically non-decreasing order-preserving binary pattern: the number of units set to 1. Follow Monotonous change, thus " If and only if ", retains the order information of the load value (as opposed to one-hot encoding with unordered information). This represents any two standardized net load values ​​of the power system to be compared; These represent the number of units p1 and p2 that are set to 1 after being encoded by the thermometer.

[0029] In this step, through the above thermometer encoding, the order relationship of the load values ​​is completely preserved by the number of units set to 1, so that the binary visible layer vector does not lose the order information of the net load while maintaining the binary structure. This provides an order-preserving target variable expression for subsequent conditional energy modeling, hence the name order-preserving encoding.

[0030] To determine the adjustable quantization level, this embodiment takes... , corresponding to The interval is discretized into 33 levels (quantization step size approximately...). This value matches the quantization resolution to the typical fluctuation range of the net load (on the order of 5%–10%), avoiding information loss due to overly coarse quantization while also ensuring the visible layer dimension is within acceptable limits. Keep it at a size that allows for efficient training.

[0031] Quantization series The user pre-selects the parameters based on the fluctuation range of the net load; the greater the fluctuation range, the better. A larger value provides higher quantization resolution in areas of high volatility.

[0032] II. Double-biased injection.

[0033] Figure 2 shows a data flow diagram of the method proposed in this invention.

[0034] In this step, a continuous conditional vector is constructed. It is composed of historical load windows, daily type codes, and meteorological screening features. This represents the dimension of the continuous conditional vector, i.e., the total number of features after concatenation, typically 275. The meteorological features are selected using Principal Component Analysis (PCA) and Mutual Information (MI). The first stage involves unsupervised principal component analysis for dimensionality reduction on the historical set; The second stage selects components based on the mutual information between each principal component and the net load. One principal component.

[0035] Thus, a continuous condition vector is constructed. .

[0036] Specifically, in the dual-bias conditional injection step, the continuous conditional vector u changes the conditional energy function only by modulating the visible and hidden biases (i.e., it does not interact multiplicatively with the activation states of the binary visible and hidden layers), and the modulation of the visible and hidden biases is achieved by two independent sets of learnable linear mapping matrices. , Implement them separately:

[0037] in, Indicates the visible bias. This indicates the hidden bias. Indicates static visible bias. This indicates static hidden bias. and This represents two sets of learnable linear mapping matrices. This represents a continuous conditional vector. This indicates the number of units in the binary visible layer, i.e., the length of the thermometer code, typically 32. This indicates the number of units in the hidden layer, typically 64.

[0038] Static visible bias The meaning is: a fundamental bias term that does not change with the continuous conditional vector u, reflecting the inherent activation tendency of each unit in the visible layer without considering external conditions; during training, it is used as a learnable parameter, along with the dynamic bias component. The addition constitutes the visible bias. .

[0039] Static hidden bias The meaning is: the basic bias term of the hidden layer that does not change with the continuous conditional vector u, reflecting the inherent activation tendency of the hidden unit; and the dynamic bias component. Adding them together creates a hidden bias. .

[0040] Corresponding conditional energy function:

[0041] in, Represents the conditional energy function; Represents the binary visible layer vector. Represents the binary hidden layer vector; Indicates the number of visible layer units. Indicates the number of hidden units; This represents the binary value of the i-th unit in the visible layer. This represents the binary value of the j-th unit in the hidden layer. Visible-hidden weights This represents the connection weight between the i-th unit of the visible layer and the j-th unit of the hidden layer; This represents the conditional bias of the i-th unit in the visible layer (i.e., the i-th component of â=a+uA). This represents the conditional bias of the j-th unit in the hidden layer (i.e., =b+uB's j-th component. Indicates the index of the visible layer cell. Indicates the index of the hidden layer unit.

[0042] Conditional distribution ( (When fixed, it is the same shape as the standard RBM)

[0043] in, This represents the conditional probability that the j-th unit of the hidden layer takes the value 1, given a binary visible layer vector v and a continuous condition vector u. Let represent the conditional probability that the i-th unit in the visible layer is 1, given a binary hidden layer vector h and a continuous conditional vector u. For the sigmoid function, This represents the real number argument of the sigmoid function (here, the net input of the cell).

[0044] because It only applies to the bias (the translation term of the energy landscape), does not enter the visible layer for value taking, and is not multiplied by any unit activation. The real order of magnitude of the continuous condition is absorbed by the bias term, while the weight... By simply adapting the interactions between binary layers, the magnitude mismatch can be fundamentally resolved.

[0045] As shown in Figure 2, the modulation of the bias and the modulation of the hidden bias are represented by two independent sets of learnable linear mapping matrices. , Implement them separately; then, continuous condition vectors Modeling is achieved by modulating the visible and hidden biases in the conditional energy function. The visible bias is then applied to the binary visible layer, and the hidden bias to the hidden layer, to adjust the activation probabilities of nodes. The weights connecting the visible and hidden layers are then used in the modeling process. This indicates the strength of the association between the two.

[0046] III. Conditional Contrast Divergence Training.

[0047] As shown in Figure 2, in this step, the conditional Boltzmann machine is trained based on the conditional free energy difference contrast divergence.

[0048] Conditional free energy Contrastive Divergence (CD-) )loss.

[0049] in, Represents conditional free energy; .

[0050] By binary hidden unit Analyze marginalization (i.e.) right Summation, Represents the conditional energy function in the above summation. Naturally obtained:

[0051] In the formula: The loss function of a conditional Boltzmann machine; Batch size (number of samples in one training iteration); positive phase For real data samples in the training batch; negative phase by Starting point Step conditional Gibbs sampling obtained (alternating execution) , common Second-rate, That is, the number of Gibbs sampling steps for the contrastive divergence (typically 1-10). This represents the visible layer update vector (intermediate state of negative phase samples) obtained by backsampling from the hidden layer during the conditional Gibbs sampling process. This represents the conditional probability distribution of the values ​​of each unit in the hidden layer, given v and u. This represents the conditional probability distribution of the values ​​of each unit in the visible layer given h and u, and is treated as a constant during backpropagation, not participating in gradient backpropagation (i.e., gradient truncation is performed on negative phase samples). This represents the continuous conditional vector corresponding to the nth sample in the training batch; It represents the conditional free energy.

[0052] Gradient joint update of all parameters No phased training is required; typical .

[0053] IV. Conditional Sampling Probability Prediction.

[0054] In this step: Determine the continuous condition vector of the target net load to be predicted, and use it as the test condition vector; Perform conditional sampling on the test condition vector; The empirical predicted distribution of the net load of the target to be predicted is obtained by decoding the sampling results, and then a continuous hierarchical probability score is obtained.

[0055] Before performing conditional sampling on the test condition vector, a preheating step is performed to bring the Markov chain to a steady-state distribution.

[0056] Specifically: As shown in Figure 2, given a test condition vector, the test condition vector is kept fixed, and a random binary vector is used as the seed for execution. First, preheat (to bring the Markov chain to a steady-state distribution and discard transitional samples), then continue execution. Subconditional Gibbs sampling Each time, the sampled binary vector is decoded into a per-unit load. Thus, the empirical predicted distribution of net load is obtained.

[0057] in, Let represent the binary visible layer vector obtained by the κ-th conditional Gibbs sampling. Indicates the sampling order index. This represents the total number of samples (the sample size of the empirically predicted distribution, typically 200). This represents the per-unit net load prediction value obtained by decoding the κ-th sample vector. This represents the binary value of the i-th visible unit in the κ-th sampling vector.

[0058] Based on this, output point prediction (sample mean) ) and Continuous Ranked Probability Score (CRPS):

[0059] In the formula, Indicates a continuous graded probability score; For the true net load, This represents the traversal index when averaging K predicted distribution samples. This represents the per-unit predicted net load value corresponding to the κ′th sample.

[0060] Continuous grading probability score is The empirical distribution formed by the samples is calculated in ensemble form, including a correction term reflecting the dispersion of the predicted distribution. Wherein: The first term is the average absolute deviation between the prediction and the true value; The second term is a correction term that reflects the dispersion of the predicted distribution. Its function is to increase the sharper the distribution and the lower the CRPS, under the premise of the same average deviation, thus encouraging the predicted distribution to be both accurate and sharp.

[0061] typical , , This represents the number of burn-in steps in a Markov chain, i.e., the number of steps taken to bring the chain to a steady-state distribution and discard transitional samples before formal sampling.

[0062] This invention can introduce high-dimensional continuous conditional features while maintaining the binary structure of the visible layer of the energy model without causing magnitude mismatch and training divergence, thereby stably outputting the probability distribution prediction of the net load.

[0063] Compared with the prior art, the technical effects of the present invention are reflected in: (1) Avoiding magnitude mismatch and training stability from the architecture - continuous information only acts on the bias rather than the binary layer. After actual training, the standard deviation of the weights is stable at about 0.23 (not tending to zero, not diverging). However, when the continuous load is directly input into the binary visible layer for training, the weight norm grows unbounded and the model cannot converge. The two form a sharp contrast, proving that the stability of the present invention comes from the architecture selection rather than parameter adjustment, which is an unexpected technical effect. (2) High-dimensional continuous meteorology—dual bias injection can be used to enable the model to absorb hundreds of dimensions of continuous exogenous features (275 dimensions in the example); (3) The probabilistic prediction effect of the three elements synergy - the order-preserving binary coding maintains the binary structure, the principal component analysis + mutual information two-stage screening provides high-quality conditional features, and the double bias injection resolves the magnitude conflict of high-dimensional continuous features entering the binary model. The three elements work together to achieve a probabilistic prediction effect of CRPS of 0.0098 (winter heavy) / 0.0092 (mid-load) / 0.0225 (summer heavy) on three typical days of a certain provincial power grid. The replacement of any link will destroy the above synergy (such as using one-hot coding will lose load sequence information, and continuous features directly entering the visible layer will cause magnitude mismatch and divergence), reflecting the synergy gain; (4) Adjustable quantization resolution – The thermometer encoding retains the sequential information of the load value, and the quantization resolution is adjustable. control( (Higher resolution means higher resolution). Users can select an appropriate resolution based on the severity of system fluctuations. .

[0064] The following detailed description uses net load probability prediction under the scenario of intraday grid congestion management as an application scenario of this invention.

[0065] This embodiment uses the IEEE RTS-96 standard single-area test system (75 buses / 145 branches / 32 units) as the verification platform. The data source is the 2023-2025 forecast data of unified dispatch load and wind and solar power output (15-minute resolution) disclosed by a provincial power grid. The meteorological data is taken from the ERA5 reanalysis dataset (21 meteorological variables, 0.25° resolution, resampled to 15 minutes). The net load is normalized to the training set peak value. .

[0066] Conditional vector composition: 168-dimensional historical load (168-hour window compressed using 4-step mean) + 7-dimensional daily type (one-hot encoding from Monday to Sunday) + Weather screening features; hour .

[0067] Two-stage screening of meteorological characteristics: Phase One in History Upper-level hourly fitting principal component analysis (PCA, unsupervised, with cumulative variance up to 90%). Phase Two Calculate each principal component (The i-th principal component) and net load Mutual Information (MI) Retain the maximum mutual information One principal component.

[0068] Two-stage fitting only uses Prior to all testing days, strict measures were taken to prevent data leakage.

[0069] In this embodiment, PCA is implemented using standard principal component decomposition, and mutual information is based on... The nearest neighbor estimator implementation ( In the absence of the aforementioned toolkits, PCA can be equivalently implemented by performing singular value decomposition (SVD) on the centered feature matrix, and mutual information can be equivalently implemented by using kernel density estimation or based on... The nearest neighbor entropy estimation method is implemented without relying on any specific software package.

[0070] Network size and training: Visible layer (Quantization step size approximately) (matching typical fluctuations in net load), hidden layer Conditional dimension ( The CD-10 optimizer (learning rate 0.001, weight decay 1e-3), gradient clipping 1.0, 60 epochs, and batch size 64 were used.

[0071] reasoning: fixed Preheating step, Sub-Gibbs sampling constructs an empirical distribution. A fixed random seed (42 / 123 / 2024) can completely reproduce the distribution point-by-point. Raw results such as sampled arrays, true net load sequences, and timestamps can be archived for independent verification.

[0072] Actual test results: The standard deviation of the weights stabilized at approximately 0.23 after training (consistent across all three seasons); Figure 3 shows a bar chart of the measured continuous graded probability scores (CRPS) for typical days in the three seasons (winter peak / waisted slender ...

[0073] The CRBM of this invention outputs a complete probability distribution rather than a single-point estimate, which is something that deterministic prediction methods (such as XGBoost and LSTM) cannot provide.

[0074] The key difference between this invention and the prior art lies in: (1) Existing RBM-type methods (including several variants for tasks such as time series prediction or structured output prediction) typically use factorized ternary multiplication interactions. Introducing conditions, or directly connecting conditions to the binary visibility layer, where The condition vector represents the k-th component, where in existing methods, the conditional features are directly applied to the visible-hidden unit pair in the form of a third-order tensor multiplication. In this invention, the condition only modulates the bias, is not multiplied with any unit activation, does not enter the binary visible layer, and is represented by two sets of independent linear mappings. , By injecting visible and hidden biases respectively, the magnitude mismatch is avoided from a mechanistic perspective.

[0075] (2) Existing thermometer coding technology is used for adversarial sample defense in feedforward networks; this invention uses it as the order-preserving coding of net load target variables in the visible layer of the energy model to serve probabilistic prediction. The application objects and purposes are different.

[0076] (3) Existing net load probability prediction methods based on diffusion models employ a multi-step denoising training-sampling mechanism; this invention is based on an energy Boltzmann distribution model and employs contrastive divergence training and conditional Gibbs sampling, which are completely different technical approaches.

[0077] In summary, the inventive technical points that demonstrate the method proposed in the embodiments of the present invention are: This invention injects continuous conditional features into the conditional energy function by modulating the visible and hidden biases through only two sets of independent linear mappings. This allows the real magnitude of the continuous conditions to be absorbed by the bias term, and the weights only need to be adapted to the interaction between binary layers. This resolves the magnitude mismatch when the energy model processes continuous variables at the architectural level, enabling the energy-based probabilistic model to stably output the probability distribution prediction of the net load of the power grid.

[0078] Example 2 This embodiment discloses a power grid net load probability prediction system based on a conditional Boltzmann machine.

[0079] A grid net load probabilistic prediction system based on a conditional Boltzmann machine includes: The sequence-preserving binary encoding module is configured to: encode the standardized net load of the power system into a binary visible layer vector; The continuous condition construction module is configured to construct a continuous condition vector based on historical load data of the power system. The bias modulation module is configured to enable the continuous condition vector to change the conditional energy function in the conditional Boltzmann machine only by modulating the visible bias and the hidden bias; The training and prediction module is configured to: train the parameters of the conditional Boltzmann machine, and based on the trained conditional Boltzmann machine, obtain the empirical prediction distribution and continuous hierarchical probability score of the target net load to be predicted.

[0080] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0081] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for probabilistic prediction of grid net load based on conditional Boltzmann machines, characterized in that, Includes the following steps: The standardized net load of the power system is binary-coded into a binary visible layer vector; Constructing continuous conditional vectors based on historical load data of the power system; This allows the continuous conditional vector to change the conditional energy function in the conditional Boltzmann machine only by modulating the visible and hidden biases; The parameters of the conditional Boltzmann machine are trained, and based on the trained conditional Boltzmann machine, the empirical predicted distribution and continuous graded probability score of the target net load are obtained.

2. The grid net load probability prediction method based on conditional Boltzmann machine as described in claim 1, characterized in that, The per-unit power system net load is binary encoded into a binary visible layer vector, specifically including: The net load of the standardized power system is discretized into a binary visible layer vector using thermometer encoding; The thermometer encoding makes the binary visible layer vector present a monotonically non-increasing or monotonically non-decreasing order-preserving binary mode, so that the number of units set to 1 changes monotonically with the net load value. Quantization series satisfy ,forward One component is 1, and the rest are 0, which can be expressed by the formula: ； in, The number of units that are set to 1; This indicates the rounding operation; This represents the net load of the standardized power system; Indicates the adjustable quantization level; This represents the binary visible layer vector of the thermometer's encoded output, which is the ordered binary vector mapped after quantization, with the first k cells set to 1 and the rest set to 0.

3. The grid net load probability prediction method based on conditional Boltzmann machine as described in claim 1, characterized in that, Continuous condition characteristics include historical load windows, daily type unique thermal coding, and meteorological screening features; The meteorological screening features were obtained through a two-stage screening process: The first stage involves unsupervised principal component analysis for dimensionality reduction on the historical set; The second stage selects components based on the mutual information between each principal component and the net load. One principal component.

4. The grid net load probability prediction method based on conditional Boltzmann machine as described in claim 1, characterized in that, The continuous conditional vector participates in modeling only by modulating the visible and hidden biases in the conditional energy function, without interacting with the activation states of the binary visible layer or the activation states of the hidden layer through multiplication. The modulation of the visible bias and the modulation of the hidden bias are represented by two independent sets of learnable linear mapping matrices. , Implement them separately: ； ； in, Indicates the visible bias. This indicates the hidden bias. Indicates static visible bias. This indicates static hidden bias. and This represents two sets of learnable linear mapping matrices. Represents a continuous conditional vector; Static visible bias The basic bias term is independent of the continuous condition vector u, reflecting the inherent activation tendency of each unit in the visible layer when external conditions are not considered. Static hidden bias This is the basic bias term of the hidden layer that does not change with the continuous conditional vector u, reflecting the inherent activation tendency of the hidden unit.

5. The grid net load probability prediction method based on conditional Boltzmann machine as described in claim 4, characterized in that, The conditional energy function is expressed as: ； in, Represents the conditional energy function; Represents the binary visible layer vector. Represents the binary hidden layer vector; Indicates the number of visible layer units. Indicates the number of hidden units; This represents the binary value of the i-th unit in the visible layer. This represents the binary value of the j-th unit in the hidden layer. This represents the connection weight between the i-th unit of the visible layer and the j-th unit of the hidden layer; This represents the conditional bias of the i-th unit in the visible layer. This represents the conditional bias of the j-th unit in the hidden layer; Indicates the index of the visible layer cell. Indicates the index of the hidden layer unit.

6. The grid net load probability prediction method based on conditional Boltzmann machine as described in claim 1, characterized in that, The training process of the conditional Boltzmann machine adopts the contrastive divergence loss based on the difference of conditional free energy. The positive phase is the real data sample, and the negative phase sample is obtained by several steps of conditional Gibbs sampling and gradient truncation is performed. All parameters are updated jointly.

7. The grid net load probability prediction method based on conditional Boltzmann machine as described in claim 6, characterized in that, Conditional free energy can be expressed as: ； in, Represents conditional free energy; This represents the conditional bias of the i-th unit in the visible layer. This represents the binary value of the i-th unit in the visible layer. , The real-valued argument of the softplus function; This represents the conditional bias of the j-th unit in the hidden layer; This represents the connection weight between the i-th unit of the visible layer and the j-th unit of the hidden layer; Loss function of conditional Boltzmann machine for: ； in, For batch size, positive phase For real data samples in the training batch; negative phase by Starting point Conditional Gibbs sampling was obtained; This represents the continuous conditional vector corresponding to the nth sample in the training batch; It represents the conditional free energy.

8. The grid net load probability prediction method based on conditional Boltzmann machine as described in claim 1, characterized in that, Also includes: Determine the continuous condition vector of the target net load to be predicted, and use it as the test condition vector; Perform conditional sampling on the test condition vector; The empirical predicted distribution of the net load of the target to be predicted is obtained by decoding the sampling results, and then a continuous hierarchical probability score is obtained.

9. The grid net load probability prediction method based on conditional Boltzmann machine as described in claim 8, characterized in that, Before performing conditional sampling on the test condition vector, a preheating step is performed to make the Markov chain reach a steady-state distribution. The specific formula for continuous graded probability scoring is as follows: ； in, This represents a continuous graded probability score. Indicates the total number of samples; This represents the per-unit net load prediction value obtained by decoding the κ-th sample vector; This represents the true net load. This represents the traversal index when averaging K predicted distribution samples; κ represents the per-unit net load prediction value corresponding to the κ′th sample; κ represents the sampling order index, ranging from 1 to K.

10. A grid net load probability prediction system based on a conditional Boltzmann machine, characterized in that, include: The sequence-preserving binary encoding module is configured to: encode the standardized net load of the power system into a binary visible layer vector; The continuous condition construction module is configured to construct a continuous condition vector based on historical load data of the power system. The bias modulation module is configured to enable the continuous condition vector to change the conditional energy function in the conditional Boltzmann machine only by modulating the visible bias and the hidden bias; The training and prediction module is configured to: train the parameters of the conditional Boltzmann machine, and based on the trained conditional Boltzmann machine, obtain the empirical prediction distribution and continuous hierarchical probability score of the target net load to be predicted.

Images