Non-parametric multi-element probability prediction method and system for new energy power system based on monotone neural network

Abstract

The application discloses a new energy power system non-parametric multivariate probability prediction method and system based on monotone neural network. The application firstly acquires a historical time sequence characteristic vector; constructs an autoregressive model composed of multiple cascaded constraint monotone neural network submodules, each submodule outputs a cumulative distribution function value and is monotone non-decreasing relative to an input target variable component; during training, the real value of the target variable component is taken as input, the probability density function value is calculated by using automatic differentiation, and the model is trained by taking negative log likelihood as loss; during prediction, multiple sampling values of a multivariate joint distribution are generated by inverse transformation sampling to obtain a probability prediction result. The application fully utilizes the strong fitting capability of the constraint monotone neural network, solves the problem that traditional multivariate probability prediction methods are difficult to simultaneously consider mathematical constraints and distribution flexibility through an autoregressive structure and a mixed activation strategy, and realizes high-precision non-parametric multivariate probability prediction of a new energy power system.

Description

Technical Field

[0001] This invention belongs to the field of power system renewable energy power prediction, and relates to a multivariate probability prediction method and system, particularly a nonparametric multivariate probability prediction method and system for renewable energy power systems based on monotonic neural networks. Background Technology

[0002] The penetration rate of new energy sources, represented by wind power and solar power, in the power system is rapidly increasing. However, the generation of new energy power is highly random, volatile, and intermittent, and its large-scale grid connection poses a significant challenge to the safe and stable operation of the power system and real-time power balance. Accurate power forecasting is key to solving this problem.

[0003] Current forecasting methods are mainly divided into point forecasting and probabilistic forecasting. Traditional point forecasting can only provide a single forecast value and cannot quantify the uncertainty of the forecast result, making it difficult to meet the needs of power system risk assessment, reserve capacity optimization, and economic dispatch. Therefore, probabilistic forecasting techniques that can provide forecast intervals or probability density functions have gradually become a research hotspot.

[0004] Among existing probabilistic prediction techniques, parametric methods are computationally simple, but new energy power data often exhibit skewed, peaked, or multi-peaked distributions, and the assumption of a fixed distribution often leads to a significant decrease in prediction accuracy. While non-parametric methods avoid distribution assumptions, they still have three significant limitations in practical applications: First, monotonicity constraints are difficult to guarantee; the cumulative distribution function must mathematically satisfy the property of being monotonically non-decreasing. Traditional quantile regression models model different quantiles separately, which easily leads to quantile crossover, violating mathematical principles and making the prediction results physically uninterpretable. Second, modeling multivariate dependencies is difficult; new energy power plants often exist in clusters, with complex spatiotemporal correlations between different plants and between different times within the same plant. Existing methods often use univariate independent prediction, ignoring the coupling relationships between variables, resulting in an inability to accurately assess the overall risk of the system. Third, model fitting ability is limited; ordinary neural networks lack structural monotonicity constraints when fitting the cumulative distribution function. Forcibly adding penalty terms during training not only makes convergence difficult but also makes it difficult to strictly guarantee monotonicity across the entire domain. Furthermore, simple monotonic network structures often lack sufficient expressive power to fit complex non-Gaussian distribution shapes.

[0005] Therefore, there is an urgent need for a non-parametric probabilistic prediction method that can strictly guarantee mathematical constraints, has strong nonlinear fitting capabilities, and can effectively capture the spatiotemporal dependencies of multivariate variables, so as to achieve high-precision modeling of new energy power systems. Summary of the Invention

[0006] To address the shortcomings of existing methods for predicting the probabilistic distribution of renewable energy power, such as insufficient fitting ability when handling complex distributions, difficulty in ensuring the monotonicity constraint of the cumulative distribution function, and neglect of dependencies between multivariate variables, this invention aims to provide a nonparametric multivariate probabilistic prediction method and system for renewable energy power systems based on monotonic neural networks. This invention achieves high-precision, nonparametric modeling of the joint probability distribution of multiple variables in renewable energy power systems by constructing a constrained monotonic neural network containing hybrid activation functions and hybrid monotonic linear layers, combined with autoregressive mechanisms and automatic differentiation techniques.

[0007] To achieve the above objectives, the present invention proposes the following technical solution:

[0008] A nonparametric multivariate probabilistic prediction method for new energy power systems based on monotonic neural networks includes the following steps:

[0009] Historical observation data of the new energy power system is acquired, and historical time-series feature vectors are extracted using a time-series encoder.

[0010] A multivariate joint probability prediction model based on an autoregressive structure is constructed. The multivariate joint probability prediction model includes multiple constrained monotonic neural network sub-modules connected in an autoregressive order. Each constrained monotonic neural network sub-module corresponds to a target variable component in the vector of the target variable to be predicted.

[0011] For the i-th constrained monotonic neural network submodule, the historical time series feature vector and the first i-1 target variable components are used as context condition inputs, and the i-th target variable component is used as independent variable input. The conditional cumulative distribution function value of the i-th target variable component under the context condition is output; wherein, the conditional cumulative distribution function value is monotonically non-decreasing relative to the i-th target variable component.

[0012] During the model training phase, the true values ​​of each target variable component are input into the corresponding constrained monotonic neural network submodule, and the true values ​​of the first i-1 target variable components are used as part of the context conditions of the i-th constrained monotonic neural network submodule. The corresponding conditional probability density function value is calculated based on the conditional cumulative distribution function value output by each constrained monotonic neural network submodule, and the model parameters are trained using the negative log-likelihood function as the loss function.

[0013] In the model prediction stage, based on the trained multivariate joint probability prediction model, the sampling values ​​of each target variable component are generated sequentially using the inverse transformation sampling method in the autoregressive order. When generating the sampling value of the i-th target variable component, the historical time series feature vector and the sampling values ​​of the first i-1 target variable components are used as context conditions to obtain multiple prediction samples that conform to the multivariate joint distribution. The multiple prediction samples are then used to obtain the probabilistic prediction results of the new energy power system.

[0014] Furthermore, the constrained monotonic neural network submodule includes a hybrid linear monotonic layer and a hybrid activation function layer; the hybrid linear monotonic layer is used to perform a linear transformation on the input features and ensure that the cumulative distribution function value is monotonically non-decreasing relative to the target variable component; the hybrid activation function layer is used to perform a nonlinear mapping on the linearly transformed features, and enhances the model's ability to fit complex non-Gaussian distribution shapes by combining the characteristics of multiple activation functions.

[0015] Furthermore, the hybrid linear monotonic layer divides the input weights into monotonic weights corresponding to the independent variable input and non-monotonic weights corresponding to the context condition input, and forces the monotonic weights to be non-negative through the Softplus function, thereby ensuring the monotonic non-decreasing property of the linear transformation with respect to the target variable.

[0016] Furthermore, the hybrid activation function layer employs a piecewise strategy to perform nonlinear mapping on the features after linear transformation. Specifically, the features output by the hybrid linear monotonic layer are divided into a convex function part, a concave function part, and a bounded part along the channel dimension. The convex function part uses the LeakyReLU activation function to fit the left tail characteristics of the distribution; the concave function part uses the reflection LeakyReLU activation function to fit the saturation region characteristics of the distribution; and the bounded part uses the Tanh activation function to achieve a smooth transition. The three parts of features after the above nonlinear mapping are concatenated along the channel dimension as the output of the hybrid activation function layer, thereby constructing a monotonic activation function that can flexibly fit any cumulative distribution function shape.

[0017] Furthermore, the multivariate joint probability prediction model based on autoregressive structure decomposes the multivariate joint probability density function into a product of a series of conditional probability density functions. Each constrained monotonic neural network submodule is used to fit the cumulative distribution function of the corresponding target variable component, and the conditional probability density function value is obtained by differentiating the cumulative distribution function.

[0018] Furthermore, the conditional probability density function value is obtained by taking the derivative of the cumulative distribution function, which is achieved by calculating the partial derivative of the cumulative distribution function value with respect to the input target variable component using the automatic differentiation mechanism of the deep learning framework.

[0019] Furthermore, the loss function is the sum of the negative log-likelihoods of all target variables.

[0020] Furthermore, the step of generating sampled values ​​for each target variable component sequentially using the inverse transform sampling method according to the autoregressive order specifically includes:

[0021] Generate random numbers that follow a uniform distribution (0,1) with the same number of components as the target variable, and form a random vector;

[0022] According to the autoregressive order of the target variable components, for the i-th target variable component, the historical time series feature vector and the sampled values ​​of the first i-1 target variable components are fixed, and the i-th target variable component is used as the independent variable to perform a binary search within the preset numerical range.

[0023] The candidate values ​​obtained during the search process are input into the corresponding constrained monotonic neural network submodule to calculate the cumulative distribution function value. The binary search interval is updated according to the relationship between the cumulative distribution function value and the i-th random number in the random vector until the preset convergence condition is met. The candidate values ​​that meet the preset convergence condition are used as the predicted sample values ​​of the i-th target variable component.

[0024] By repeatedly generating multiple random vectors and performing the above operations, multiple sets of sampled values ​​describing the future state of the new energy power system are obtained, which serve as multiple prediction samples conforming to a multivariate joint distribution.

[0025] Furthermore, the historical time-series feature vector is obtained by encoding historical observation data by a time-series encoder; the time-series encoder includes a linear embedding layer, a position encoding layer, and a Transformer encoder layer.

[0026] This invention also provides a nonparametric multivariate probabilistic prediction system for new energy power systems based on monotonic neural networks, used in the above method, comprising:

[0027] The time-series coding module is used to acquire historical observation data of the new energy power system and extract historical time-series feature vectors through the time-series encoder.

[0028] The model building module is used to construct a multivariate joint probability prediction model based on an autoregressive structure. The multivariate joint probability prediction model includes multiple constrained monotonic neural network sub-modules connected in an autoregressive order. Each constrained monotonic neural network sub-module corresponds to a target variable component in the vector of target variables to be predicted. For the i-th constrained monotonic neural network sub-module, the historical time-series feature vector and the first i-1 target variable components are used as contextual condition inputs, and the i-th target variable component is used as the independent variable input. The module outputs the conditional cumulative distribution function value of the i-th target variable component under the contextual conditions. The conditional cumulative distribution function value is monotonically non-decreasing relative to the i-th target variable component.

[0029] The training module is used during the model training phase to input the true values ​​of each target variable component into the corresponding constrained monotonic neural network submodule, and to use the true values ​​of the first i-1 target variable components as part of the context conditions of the i-th constrained monotonic neural network submodule. Based on the conditional cumulative distribution function value output by each constrained monotonic neural network submodule, the corresponding conditional probability density function value is calculated, and the negative log-likelihood function is used as the loss function to train the model parameters.

[0030] The prediction module is used in the model prediction stage to generate sampled values ​​of each target variable component sequentially using the inverse transformation sampling method according to the autoregressive order, based on the trained multivariate joint probability prediction model. When generating the sampled value of the i-th target variable component, the historical time series feature vector and the sampled values ​​of the first i-1 target variable components are used as context conditions to obtain multiple prediction samples that conform to the multivariate joint distribution. The multiple prediction samples are then used to obtain the probabilistic prediction result of the new energy power system.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] First, by combining monotonic linear layers and Softplus weight constraints, this invention rigorously guarantees the monotonic non-decreasing nature of the cumulative distribution function of the output relative to the target variable components from the physical level of the network structure. This completely solves the quantile crossover problem commonly found in traditional non-parametric methods, ensuring that the prediction results conform to mathematical principles and have clear physical interpretability.

[0033] Secondly, this invention designs a hybrid activation function layer, which integrates the characteristics of convex, concave and bounded functions, enabling the neural network to flexibly fit complex non-Gaussian distribution shapes such as skewness, spikes and thick tails commonly found in new energy data, thus significantly improving the accuracy and reliability of probability prediction.

[0034] Furthermore, this invention employs an autoregressive structure to decompose the multivariate joint distribution into a product of conditional probabilities, which can effectively capture the complex nonlinear spatiotemporal dependencies between different renewable energy power plants or between different variables of the same power plant. Compared with independent univariate prediction, this method can provide multivariate joint probability prediction, providing a more valuable reference for the coordinated scheduling and risk assessment of power systems.

[0035] Finally, this invention utilizes automatic differentiation technology to directly optimize the log-likelihood function, realizing an end-to-end training process that is more stable. At the same time, combined with an efficient binary inverse transform sampling strategy, it can quickly generate a large number of prediction scenario samples that conform to the joint distribution, meeting the needs of real-time applications in power systems. Attached Figure Description

[0036] Figure 1This is a flowchart of the nonparametric multivariate probability prediction method for new energy power systems based on monotonic neural networks in an embodiment of the present invention.

[0037] Figure 2 This is a diagram illustrating the overall architecture of the multivariate probability prediction model based on autoregressive structure in an embodiment of the present invention.

[0038] Figure 3 The diagram below shows the data processing flow and structure of the nonparametric multivariate probability prediction model in this embodiment of the invention.

[0039] Figure 4 This is a schematic diagram of the marginal probability distribution of two stations in a specific embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] Example 1

[0042] This embodiment provides a nonparametric multivariate probabilistic prediction method for new energy power systems based on monotonic neural networks, used for joint power prediction of multiple new energy power plants (such as photovoltaic power plants and wind farms) at future times. Figure 1 , 2 As shown, the method includes the following steps:

[0043] (1) Obtain the historical time series feature vector of the new energy power system.

[0044] First, historical observation data sequences of the new energy power system are obtained, and feature vectors containing deep temporal dependencies are extracted using a time-series encoder, which serve as global context conditions for subsequent probabilistic modeling.

[0045] Let the input historical observation data sequence be... ,in The length of the historical time window. Input feature dimensions, such as historical power, meteorological data, etc. First, [the following is done / is done / etc.]. Mapping to higher-dimensional space:

[0046]

[0047] in, For embedded representation, This represents a linear mapping. Subsequently, features are extracted using a temporal encoder. In this implementation, the temporal encoder adopts a Transformer encoder structure, specifically including: a linear embedding layer, a positional encoding layer, and a Transformer encoder layer, outputting a historical temporal feature vector. ,in This represents the hidden layer dimension of the encoder. This vector It contains all the historical context information needed to predict the timing.

[0048] (2) Construct a multivariate joint probability prediction model based on autoregressive structure

[0049] According to the chain rule of probability, the high-dimensional joint probability distribution is decomposed into a product of a series of one-dimensional conditional probabilities, and a corresponding constrained monotonic neural network submodule is constructed for each variable.

[0050] Let the vector of multivariate objective variables to be predicted be... ,in The dimension of the target variable (e.g., number of stations). Based on the principle of autoregressive decomposition, the joint probability density function... Represented as:

[0051]

[0052] in, Indicates the preceding One target variable component, Let be the conditional probability density function of the i-th objective variable component.

[0053] Based on the above decomposition, k cascaded constrained monotonic neural network submodules are constructed, denoted as . Each submodule corresponds to a target variable component, used to fit its conditional cumulative distribution function. For the i-th submodule... Its input vector From global context Preceding variables and current target variable components It is pieced together:

[0054]

[0055] Among them, historical time series feature vector and the first i-1 target variable components As contextual input As an input variable.

[0056] (3) Specific design of the constrained monotonic neural network submodule

[0057] The constrained monotonic neural network submodule, by mixing monotonic linear layers and mixed activation function layers, ensures that the output is consistent with the target variable components. While being monotonically non-decreasing, it achieves flexible fitting of complex non-Gaussian distributions.

[0058] Each constrained monotonic neural network submodule contains several stacked linear-activation layers, with the last layer outputting a scalar value, which is then mapped to a sigmoid function. The interval is the cumulative distribution function value. .

[0059] Define the linear transformation of the l-th layer as To ensure that the network output is relative to the target variable components. The monotonicity must be guaranteed from the input. All weights on the output path are non-negative.

[0060] weight matrix Divided into two parts: corresponding to monotonic input (i.e., originating from...) The weight subset of the components and corresponding to conditional input (i.e., originating from) and The weight subset of the components .

[0061] right Apply Softplus constraints, while Keep unconstrained:

[0062]

[0063] The final linear transformation is calculated as follows:

[0064]

[0065] This design physically forces the network to... The partial derivatives are always non-negative.

[0066] To enhance the network's ability to fit cumulative distribution functions of arbitrary shapes, a piecewise hybrid activation function is defined. This function incorporates the output features of the linear layer. It is divided into three parts along the channel dimension: , and .

[0067] Different nonlinear transformations are applied to these three parts respectively:

[0068]

[0069]

[0070]

[0071] in, It maintains convexity. Concavity is constructed using LeakyReLU reflection. The hyperbolic tangent function is used to provide a smooth bounded transition.

[0072] Finally, the three parts are concatenated as the activated output:

[0073]

[0074] (4) Model training

[0075] The probability density function is calculated using automatic differentiation techniques, and the model is trained end-to-end based on the principle of maximum likelihood estimation.

[0076] For the i-th objective variable component, the output of the constrained monotonic neural network submodule is the conditional cumulative distribution function value. According to the definition of probability, the probability density function for right The derivative of .

[0077] The gradient is directly calculated using the automatic differentiation property of deep learning computational graphs:

[0078]

[0079] For numerical stability, its logarithmic form is usually calculated. The total loss function for the entire multivariate joint model is defined. The sum of the negative log-likelihoods of all target variables:

[0080]

[0081] in, For training batch size, For all trainable parameters of the model, To prevent logarithmic calculations from overflowing to the minimum value.

[0082] After obtaining the loss function value, the specific steps for optimizing the model parameters using gradient descent and early stopping mechanisms are as follows:

[0083] Step (1): Given the training dataset Validate dataset Initialize the model's trainable parameters Adam optimizer Learning rate Maximum number of training rounds Stop early and be patient. ;

[0084] Step (2): Set the initial training rounds Early Stop Counter Optimal verification loss ;

[0085] Step (3): Traverse the training dataset All data batches Calculate the probability density function values ​​of each variable through forward propagation, and then calculate the loss. Calculate the gradient Through the optimizer Update parameters ;

[0086] Step (4): Traverse the validation dataset All data batches Calculate the forward propagation and obtain the average verification loss. ;

[0087] Step (5): If ,but Save the current optimal model parameters Reset the early stop counter k=0; if Then the early stop counter k = k + 1;

[0088] Step (6): Training rounds

[0089] Step (7): If or If the result is positive, training is terminated; otherwise, return to step (3) to continue the next iteration.

[0090] Step (8): Training ends, return the saved optimal model parameters. This will give us a well-trained multivariate joint probability prediction model based on a monotonic neural network.

[0091] (5) In the prediction stage, an inverse transformation sampling strategy is used to generate prediction samples that conform to the joint distribution.

[0092] Since constrained monotonic neural networks fit the cumulative distribution function, and their analytical inverse function cannot be directly written, a numerical iteration method is used.

[0093] First, generate the obedience. Uniformly distributed random vector .

[0094] According to the autoregressive order of the target variable components arrive Sampling is performed sequentially:

[0095] For the i-th target variable component, the context feature vector is fixed. and the generated sample values Construct a single-variable equation:

[0096]

[0097] A binary search method is used to iteratively search within a preset numerical interval. This preset numerical interval is determined based on the physical range of the target variable or the normalized range of the training data. Specifically, when the target variable is represented in the actual power space, the numerical interval is determined by the physical lower and upper limits of the new energy output. When the target variable has undergone normalization or standardization, the numerical interval is determined by the minimum and maximum values ​​of the normalized or standardized target variable in the training set, and preset margins can be set at both ends. The iterative search continues until the error between the cumulative distribution function output value and the corresponding uniform random number satisfies a preset convergence condition; the resulting solution is the predicted sampled value. .

[0098] By repeating the above process multiple times, multiple sets of multivariate scenario samples describing the future state of the new energy power system can be generated, and the final multivariate joint probability prediction result can be obtained.

[0099] The nonparametric multivariate probability prediction method of this invention introduces a hybrid monotonic linear layer and a piecewise hybrid activation function, transforming the complex probability cumulative distribution fitting process into a deep learning optimization problem that strictly satisfies the monotonicity mathematical constraint. Furthermore, it employs an autoregressive architecture combined with automatic differentiation and inverse transformation sampling techniques to effectively capture the nonlinear spatiotemporal coupling relationship between multiple variables. This greatly overcomes the limitations of traditional parametric and nonparametric methods in terms of distribution assumptions and multidimensional dependency handling. With high-precision multivariate probability density distribution and uncertainty interval estimation, it efficiently supports risk assessment, reserve capacity optimization, and safe and economical dispatch of high-proportion renewable energy power systems.

[0100] like Figure 3 As shown, in one specific embodiment of the present invention, multivariate probability prediction is performed using photovoltaic power output data from multiple power plants in a certain region, with a data time resolution of 15 minutes. The training set, validation set, and test set of the dataset are divided in a ratio of 7:1:2. The prediction objective is to predict the joint probability distribution of multiple power plants at future times based on historical time-series data.

[0101] Taking the joint probabilistic prediction task of 5 / 10 / 15 new energy power plants as an example, the length of the input historical sequence is set to 8. The specific parameter settings and hierarchical connection relationships of the nonparametric multivariate probabilistic prediction model based on monotonic neural networks are as follows:

[0102] 1. Data Preprocessing and Feature Extraction Stage: First, the original time-series data is standardized to construct a sliding window sample set. The batch size is set to 64. The standardized multivariate sequences are input into the temporal encoder. The temporal encoder consists of a linear embedding layer, a positional encoding layer, and a single Transformer encoder layer, where the Transformer multi-head attention head is set to 2, and the feedforward network dimension is set to 64. After the historical sequences pass through the encoder, the output of the last time step is extracted to form a global context feature vector with a dimension of 32. At this point, the tensor dimension of the input data is transformed from (64, 8, 5 / 10 / 15) to a feature tensor (64, 32) containing global context information.

[0103] 2. Autoregressive Joint Modeling Layer: In order to capture the spatiotemporal dependencies among 5 / 10 / 15 stations, the model constructs 5 / 10 / 15 cascaded constrained monotonic neural network sub-modules.

[0104] 1) For station 1: The input dimension of its constrained monotonic neural network submodule is 33, which is composed of global context features of dimension 32 and station 1 target power variables of dimension 1.

[0105] 2) For station i between station 2 and 5 / 10 / 15: the input dimension of its constrained monotonic neural network submodule is 32+i, which is composed of global context features of dimension 32, the previous true value of station i-1 of dimension i, and the target power variable component of station i of dimension 1, thereby realizing the chain propagation of conditional probability.

[0106] 3. Constrained Monotonic Neural Network Core Module: Each constrained monotonic neural network submodule has a hidden layer dimension of 60 and contains three layers of hybrid monotonic linear layers and hybrid activation function layers, aiming to map input features to the cumulative distribution space. The specific configuration of this embodiment is as follows:

[0107] 1) Input layer to hidden layer: The first layer, a hybrid monotonic linear layer, maps the input (dimension 33 or 34) to a hidden dimension of 60. The network forces the weights of the input corresponding to the target power variable component to be non-negative through the Softplus function, while the weights of the corresponding context and preceding variables remain unconstrained.

[0108] 2) Hybrid Activation Layer: This layer divides the 60-dimensional features output from the hidden layer into three parallel processing units (each unit has a dimension of 20) along the channel dimension. The first unit is activated by LeakyReLU with a negative slope of 0.01 to fit the convex regions of the distribution; the second unit is activated by reflection LeakyReLU to fit the concave regions of the distribution; and the third unit is activated by Tanh to provide bounded smoothness. These three activation vectors are fused by concatenation, outputting a tensor of shape (64, 60), which serves as the input to subsequent networks.

[0109] 3) Hidden layer mapping to output: Subsequent mixed monotonic linear layers constrain all input weights to be non-negative, ensuring that the network is strictly monotonic and non-decreasing.

[0110] 4. Probability Output and Sampling Stage: The one-dimensional output of the constrained monotonic neural network is mapped to the (0, 1) interval through the Sigmoid activation function to obtain the corresponding conditional cumulative distribution function value.

[0111] During training, the automatic differentiation mechanism of the deep learning framework is used to differentiate the cumulative distribution function (CDF) value, and the probability density function (PDF) value with tensor dimension (64, 5 / 10 / 15) is calculated for the calculation of negative log-likelihood loss.

[0112] In the prediction phase, a binary search method is used to iteratively approximate within a set standardized numerical range (e.g., [-5, 5]), and the output tensor dimension is reshaped into a multivariate joint prediction sampling matrix (64, 300, 5 / 10 / 15) (where 300 is the number of generated random prediction scenarios), thereby constructing a high-precision nonparametric empirical multivariate probability distribution.

[0113] Table 1 presents the prediction performance indices obtained by the method of the present invention and the comparative method, wherein the comparative method uses DeepAR for prediction.

[0114]

[0115] As shown in Table 1, under different test conditions with 5, 10, and 15 testing sites, the method of this invention achieves superior results compared to the comparative method in two key performance indicators: Average Absolute Proportion Deviation (AAPD) and Multivariate Continuous Probability Ranking Score (sum_CRPS). The former reflects the reliability of the prediction model, while the latter reflects the overall performance of multivariate probability prediction. This indicates that the method of this invention has higher reliability and prediction performance, and it maintains a stable performance advantage even with an increase in the number of testing sites, verifying the superiority of the method of this invention in multivariate joint probability prediction problems.

[0116] Figure 4 The graph shows the probability density function of two photovoltaic power plants under joint probability prediction. As can be seen from the graph, the multivariate probability prediction method provided by this invention can capture the spatiotemporal coupling relationship between different photovoltaic power plants. It should be noted that, in addition to multi-plant photovoltaic power probability prediction, the nonparametric multivariate probability prediction method for new energy power systems based on monotonic neural networks proposed in this invention is also applicable to other new energy multivariate probability prediction scenarios in power systems, such as multi-plant wind power probability prediction and multi-regional load probability prediction, demonstrating broad applicability.

[0117] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0118] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0121] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall still fall within the protection scope of the technical solutions of the present invention.

Claims

1. A nonparametric multivariate probabilistic prediction method for new energy power systems based on monotonic neural networks, characterized in that, Includes the following steps: Historical observation data of the new energy power system is acquired, and historical time-series feature vectors are extracted using a time-series encoder. A multivariate joint probability prediction model based on an autoregressive structure is constructed. The multivariate joint probability prediction model includes multiple constrained monotonic neural network sub-modules connected in an autoregressive order. Each constrained monotonic neural network sub-module corresponds to a target variable component in the vector of the target variable to be predicted. For the i-th constrained monotonic neural network submodule, the historical time series feature vector and the first i-1 target variable components are used as context condition inputs, and the i-th target variable component is used as independent variable input. The conditional cumulative distribution function value of the i-th target variable component under the context condition is output; wherein, the conditional cumulative distribution function value is monotonically non-decreasing relative to the i-th target variable component. During the model training phase, the true values ​​of each target variable component are input into the corresponding constrained monotonic neural network submodule, and the true values ​​of the first i-1 target variable components are used as part of the context conditions of the i-th constrained monotonic neural network submodule. The corresponding conditional probability density function value is calculated based on the conditional cumulative distribution function value output by each constrained monotonic neural network submodule, and the model parameters are trained using the negative log-likelihood function as the loss function. In the model prediction stage, based on the trained multivariate joint probability prediction model, the sampling values ​​of each target variable component are generated sequentially using the inverse transformation sampling method in the autoregressive order. When generating the sampling value of the i-th target variable component, the historical time series feature vector and the sampling values ​​of the first i-1 target variable components are used as context conditions to obtain multiple prediction samples that conform to the multivariate joint distribution. The multiple prediction samples are then used to obtain the probabilistic prediction results of the new energy power system.

2. The nonparametric multivariate probabilistic prediction method for new energy power systems according to claim 1, characterized in that, The constrained monotonic neural network submodule includes a hybrid linear monotonic layer and a hybrid activation function layer; the hybrid linear monotonic layer is used to ensure the monotonicity and non-decreasingness of the cumulative distribution function value with respect to the target variable components; the hybrid activation function layer is used to enhance the model's ability to fit complex non-Gaussian distribution shapes.

3. The nonparametric multivariate probabilistic prediction method for new energy power systems according to claim 2, characterized in that, The hybrid linear monotonic layer divides the input weights into monotonic weights corresponding to the independent variable inputs and non-monotonic weights corresponding to the context condition inputs, and forces the monotonic weights to be non-negative values ​​through the Softplus function.

4. The nonparametric multivariate probabilistic prediction method for new energy power systems according to claim 2, characterized in that, The hybrid activation function layer divides the features output by the hybrid linear monotonic layer into a convex function part, a concave function part, and a bounded part in the channel dimension. The LeakyReLU activation function, the reflection LeakyReLU activation function, and the Tanh activation function are used for nonlinear mapping, and the three mapped parts are concatenated as the output.

5. The nonparametric multivariate probabilistic prediction method for new energy power systems according to claim 1, characterized in that, The multivariate joint probability prediction model based on autoregressive structure decomposes the multivariate joint probability density function into a product of a series of conditional probability density functions. Each constrained monotonic neural network submodule is used to fit the cumulative distribution function of the corresponding target variable component. The value of the conditional probability density function is obtained by taking the derivative of the cumulative distribution function.

6. The nonparametric multivariate probabilistic prediction method for new energy power systems according to claim 5, characterized in that, The conditional probability density function value is obtained by taking the derivative of the cumulative distribution function, which is achieved by using the automatic differentiation mechanism of the deep learning framework to calculate the partial derivative of the cumulative distribution function value with respect to the input target variable components.

7. The nonparametric multivariate probabilistic prediction method for new energy power systems according to claim 1, characterized in that, The loss function is the sum of the negative log-likelihoods of all target variables.

8. The nonparametric multivariate probabilistic prediction method for new energy power systems according to claim 1, characterized in that, The step of generating sampled values ​​for each target variable component sequentially using the inverse transform sampling method according to the autoregressive order specifically includes: Generate random numbers that follow a uniform distribution (0,1) with the same number of components as the target variable, and form a random vector; According to the autoregressive order of the target variable components, for the i-th target variable component, the historical time series feature vector and the sampled values ​​of the first i-1 target variable components are fixed, and the i-th target variable component is used as the independent variable to perform a binary search within the preset numerical range. The candidate values ​​obtained during the search process are input into the corresponding constrained monotonic neural network submodule to calculate the cumulative distribution function value. The binary search interval is updated according to the relationship between the cumulative distribution function value and the i-th random number in the random vector until the preset convergence condition is met. The candidate values ​​that meet the preset convergence condition are used as the predicted sample values ​​of the i-th target variable component. By repeatedly generating multiple random vectors and performing the above operations, multiple sets of sampled values ​​describing the future state of the new energy power system are obtained, which serve as multiple prediction samples conforming to a multivariate joint distribution.

9. The nonparametric multivariate probabilistic prediction method for new energy power systems according to claim 1, characterized in that, The historical time-series feature vector is obtained by encoding historical observation data by a time-series encoder; the time-series encoder includes a linear embedding layer, a position encoding layer, and a Transformer encoder layer.

10. A nonparametric multivariate probabilistic prediction system for new energy power systems based on monotonic neural networks, used to implement the method according to any one of claims 1-9, characterized in that, include: The time-series coding module is used to acquire historical observation data of the new energy power system and extract historical time-series feature vectors through the time-series encoder. The model building module is used to construct a multivariate joint probability prediction model based on an autoregressive structure. The multivariate joint probability prediction model includes multiple constrained monotonic neural network sub-modules connected in an autoregressive order. Each constrained monotonic neural network sub-module corresponds to a target variable component in the vector of target variables to be predicted. For the i-th constrained monotonic neural network sub-module, the historical time-series feature vector and the first i-1 target variable components are used as contextual condition inputs, and the i-th target variable component is used as the independent variable input. The module outputs the conditional cumulative distribution function value of the i-th target variable component under the contextual conditions. The conditional cumulative distribution function value is monotonically non-decreasing relative to the i-th target variable component. The training module is used during the model training phase to input the true values ​​of each target variable component into the corresponding constrained monotonic neural network submodule, and to use the true values ​​of the first i-1 target variable components as part of the context conditions of the i-th constrained monotonic neural network submodule. Based on the conditional cumulative distribution function value output by each constrained monotonic neural network submodule, the corresponding conditional probability density function value is calculated, and the negative log-likelihood function is used as the loss function to train the model parameters. The prediction module is used in the model prediction stage to generate sampled values ​​of each target variable component sequentially using the inverse transformation sampling method according to the autoregressive order, based on the trained multivariate joint probability prediction model. When generating the sampled value of the i-th target variable component, the historical time series feature vector and the sampled values ​​of the first i-1 target variable components are used as context conditions to obtain multiple prediction samples that conform to the multivariate joint distribution. The multiple prediction samples are then used to obtain the probabilistic prediction result of the new energy power system.

Images