An industry chain association-based industry electricity consumption prediction method, system and device

Abstract

The application discloses an industry electricity consumption prediction method, system and equipment based on an industry chain association, relates to the technical field of power system load prediction, and comprises the following steps: obtaining electricity consumption sequences of upstream and downstream industries, and respectively constructing upstream and downstream phase space trajectories through delay embedding; a differential homeomorphism mapping relationship between the upstream and downstream phase space trajectories is established, a joint phase space is constructed according to the mapping relationship; sustained homology analysis is performed on the joint phase space, a filter parameter value at which a homology feature suddenly changes is identified as a time-varying division point; the joint phase space is divided into multiple sub-regions according to the time-varying division point, an optimal transmission distance is calculated in each sub-region, and a dynamic time lag parameter function is obtained; the dynamic time lag parameter function is embedded into a time series neural network, network weights are jointly optimized, and a downstream industry electricity consumption prediction value is output. The application can capture continuous evolution and structural mutation of time lag, and effectively improves the prediction accuracy of industry electricity consumption under industry chain fluctuation.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and more specifically, to a method, system, and equipment for forecasting industry electricity consumption based on supply chain linkages. Background Technology

[0002] In the field of electricity consumption forecasting linked to the industrial chain, accurately capturing the transmission lag between upstream and downstream industries is crucial for achieving accurate forecasts. Because the industrial chain involves multiple stages such as raw material procurement, production and processing, and logistics, fluctuations in upstream production often require a certain time delay before being transmitted to the electricity load of downstream industries; this delay is known as the time lag effect. Traditional time lag processing methods mainly fall into two categories: one is a regression model based on a fixed lag order, which simulates industrial transmission by pre-setting a fixed time window; the other is a time lag identification method based on correlation analysis, which selects the lag order with the strongest correlation as the fixed time lag parameter.

[0003] However, these methods have significant technical flaws. The fixed time delay assumption is seriously inconsistent with the actual operating patterns of the industrial chain—during periods of raw material shortages, transmission may only take a few days; while during periods of ample supply, the transmission within the same industrial chain may take several weeks. The "electricity consumption inertia" also varies across different links in the industrial chain; the continuous production characteristics of the heavy chemical industry result in a long "signal delay" in its response to downstream demand. Although existing technologies attempt to introduce attention mechanisms, they are essentially still weighted combinations at discrete time points, unable to characterize the smooth evolution of time delays over the continuous time domain, let alone identify abrupt changes in the structural nature of time delays, leading to systematic biases in the prediction models during periods of industrial chain fluctuations.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method, system, and device for predicting industry electricity consumption based on industrial chain linkages. By constructing a joint phase space of upstream and downstream electricity consumption and identifying time-varying boundary points, the dynamic time delay parameters are calculated and then embedded into a neural network for joint optimization, thereby solving the problem that existing methods are unable to capture the nonlinear and time-varying transmission relationships between upstream and downstream of the industrial chain.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting industry electricity consumption based on supply chain linkages includes the following steps: obtaining electricity consumption sequences of upstream and downstream industries, and constructing upstream and downstream phase space trajectories respectively through delayed embedding; establishing a differential homeomorphic mapping relationship between the upstream and downstream phase space trajectories, and constructing a joint phase space based on the mapping relationship; performing continuous homology analysis on the joint phase space, and identifying the filter parameter values ​​where homology features undergo abrupt changes as time-varying boundary points; dividing the joint phase space into multiple sub-regions based on the time-varying boundary points, calculating the optimal transmission distance in each sub-region, and obtaining a dynamic time-delay parameter function; embedding the dynamic time-delay parameter function into a time-series neural network, jointly optimizing the network weights, and outputting the predicted electricity consumption value of downstream industries.

[0007] In a preferred embodiment, the construction of upstream and downstream phase space trajectories specifically involves: denoising the upstream industry electricity consumption sequence to obtain an upstream denoised sequence; using an improved CC method to jointly calculate the embedding dimension and delay time of the denoised sequence; reconstructing the phase space of the upstream denoised sequence based on the embedding dimension and delay time to obtain the upstream phase space trajectory; and repeating the above steps for the downstream industry electricity consumption sequence to obtain the downstream phase space trajectory.

[0008] In a preferred embodiment, the establishment of the differential homeomorphic mapping relationship specifically includes: calculating the local tangent space coordinates and Riemann curvature tensor of each phase point in the upstream and downstream phase space trajectories; calculating the mutual information entropy between all phase point pairs in the upstream and downstream phase space trajectories, and marking the k phase point pairs with the largest entropy values ​​as the initial corresponding anchor point set; constructing a conformal mapping distortion energy function with the local tangent space coordinates as conformal invariants, the initial corresponding anchor point set as hard constraints, and minimizing the difference in the Riemann curvature tensor as curvature consistency constraints; and solving for the differential homeomorphic mapping function that minimizes the energy function through gradient descent on the Riemann manifold.

[0009] In a preferred embodiment, forming a joint phase space specifically involves: projecting the upstream phase space trajectory onto the coordinate space of the downstream phase space trajectory using the differential homeomorphism mapping function, eliminating mapping points with abnormally high reachable density ratios; and embedding the remaining upstream projected trajectory and downstream phase space trajectory into the same coordinate space according to time index to form a joint phase space.

[0010] In a preferred embodiment, the identification of the time-varying boundary point specifically involves: transforming the trajectory points in the joint phase space into a weighted simplex and generating a filter sequence; calculating the persistent homology group on the filter sequence to obtain the persistence sequence of each homology class; performing change point detection on the persistence sequence, and taking the filter parameter value where the persistence changes abruptly as the time-varying boundary point.

[0011] In a preferred embodiment, obtaining the dynamic time delay parameter function specifically includes: dividing the region into continuous sub-regions using the time-varying boundary point as the dividing boundary; within each sub-region, obtaining the probability transition matrix of the timestamps of the upstream and downstream trajectory points by solving the entropy regularization optimal transmission, and extracting the time offset with the highest transmission probability as the local time delay constant; and performing cubic spline interpolation to smoothly connect the local time delay constants of adjacent sub-regions to obtain the dynamic time delay parameter function that changes continuously with time.

[0012] In a preferred embodiment, the step of outputting the predicted electricity consumption value for downstream industries specifically includes: embedding a dynamic time delay parameter function as the input layer of a time-series neural network as a differentiable time offset operator; extracting alignment features from the upstream historical sequence using the operator; concatenating the alignment features with the downstream historical electricity consumption data and inputting the result into the network to obtain a preliminary predicted value; constructing a joint loss function for end-to-end training; iteratively training the loss function until it converges; and then calculating the predicted electricity consumption value for downstream industries using the latest upstream electricity consumption data as input.

[0013] A system for predicting industry electricity consumption based on supply chain linkages includes: a data processing unit for acquiring electricity consumption sequences of upstream and downstream industries and constructing upstream and downstream phase space trajectories through delayed embedding; a joint phase space construction unit for establishing a differential homeomorphic mapping relationship between the upstream and downstream phase space trajectories and constructing a joint phase space based on the mapping relationship; a boundary point identification unit for performing continuous homology analysis on the joint phase space and identifying filter parameter values ​​where homology characteristics undergo abrupt changes as time-varying boundary points; a time delay calculation unit for dividing the joint phase space into multiple sub-regions based on the time-varying boundary points, calculating the optimal transmission distance in each sub-region, and obtaining a dynamic time delay parameter function; and a prediction output unit for embedding the dynamic time delay parameter function into a time-series neural network, jointly optimizing the network weights, and outputting the predicted electricity consumption value of downstream industries.

[0014] An electronic device, comprising a memory and a processor: the memory for storing a program; the processor for executing the program to implement the various steps of the industry electricity consumption forecasting method based on industrial chain linkages.

[0015] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned industry electricity consumption forecasting method based on supply chain linkages.

[0016] The technical effects and advantages of the present invention, which provides an industry electricity consumption forecasting method, system, and equipment based on industrial chain linkages, are as follows: 1. This invention obtains the electricity consumption sequences of upstream and downstream industries, constructs upstream and downstream phase space trajectories through delayed embedding, establishes a differential homeomorphic mapping relationship between the upstream and downstream phase space trajectories, and constructs a joint phase space based on this mapping relationship. This solves the problem that existing methods are difficult to nonlinearly align upstream and downstream time series data under a unified framework, enabling joint analysis of upstream and downstream electricity consumption sequences in the same coordinate space, and providing an accurate description of the industrial chain coupling state for subsequent prediction.

[0017] 2. This invention performs continuous coherence analysis on the joint phase space, identifies the filter parameter values ​​where coherence characteristics undergo abrupt changes as time-varying boundary points, divides the joint phase space into multiple sub-regions based on these boundary points, calculates the optimal transmission distance in each sub-region to obtain a dynamic time delay parameter function, embeds the dynamic time delay parameter function into a time-series neural network to jointly optimize the network weights, and outputs the predicted electricity consumption value for downstream industries. This solves the problem that existing methods are unable to capture qualitative changes in the industrial chain linkage mode and dynamic changes in transmission delay, and achieves high-precision prediction of electricity consumption for downstream industries. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an industry electricity consumption forecasting method based on supply chain linkages according to the present invention. Figure 2 This is a schematic diagram of the system structure of an industry electricity consumption forecasting method based on industrial chain linkage according to the present invention; Figure 3 A structural block diagram of an exemplary electronic device provided for implementing embodiments of the present invention; Figure 4 A schematic diagram of an exemplary storage medium that can be used to implement embodiments of the present disclosure, provided for the purposes of this invention. Figure 5 A schematic diagram illustrating the process of constructing upstream and downstream phase space trajectories provided in an embodiment of the present invention; Figure 6 This is a comparison chart of predicted and actual values ​​for the test set provided in an embodiment of the present invention. Detailed Implementation

[0019] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, Figure 1 This invention presents a method for predicting industry electricity consumption based on supply chain linkages, comprising the following steps: S1: Obtain the electricity consumption sequence of upstream and downstream industries, and construct the phase space trajectory of upstream and downstream industries through delayed embedding; like Figure 5 As shown, in this embodiment, obtaining the electricity consumption sequences of upstream and downstream industries and constructing upstream and downstream phase space trajectories through delayed embedding specifically involves: It should be noted that in this implementation, "upstream industry" and "downstream industry" are determined based on the input-output relationship of the industrial chain. Specifically, the direct consumption coefficients between each industry are calculated according to the "Input-Output Table". :

[0021] in, This represents the intermediate input amount of industry j for industry i's products. Let be the total output of industry j. When... When the threshold is exceeded (0.05 in this embodiment), industry i is defined as the upstream industry of industry j. The upstream and downstream relationship can be one-to-one, one-to-many, or many-to-one. This embodiment uses the transmission relationship between a single upstream industry and a single downstream industry as an example for explanation.

[0022] S1.1 Obtain the electricity consumption sequences of upstream and downstream industries. Specifically, this can be obtained from industry electricity load data from the power dispatch center, published industry electricity consumption statistical bulletins, or market-traded electricity data. The electricity consumption sequences are then preprocessed, including data alignment, outlier detection, and normalization.

[0023] S1.2, the upstream industry electricity consumption sequence is denoised. Specifically, singular spectrum analysis is performed on the upstream industry electricity consumption sequence to decompose it into trend terms, trend terms, and noise terms. After removing the noise terms, the upstream denoised sequence is reconstructed, as follows: Let the upstream industry electricity consumption sequence after step 0 preprocessing be: Choose a window length L that satisfies... The electricity consumption sequence of upstream industries Embedded as trajectory matrix :

[0024] in, This represents the value of the upstream industry's electricity consumption sequence at time point t. , The length of the electricity consumption sequence of upstream industries. The column number of the trajectory matrix. .

[0025] Furthermore, regarding the trajectory matrix Perform singular value decomposition:

[0026] in, The index number of the singular component. Let be the i-th eigenvalue, which is a matrix. eigenvalues, The i-th singular value reflects the energy magnitude of the i-th component. Let i be the i-th left singular vector. Let i be the i-th right singular vector. This represents the transpose of the right singular vector.

[0027] It should be noted that the specific implementation method of Singular Value Decomposition (SVD) is well known to those skilled in the art, and will not be described in detail here.

[0028] Furthermore, the L singular components are divided into three groups: trend term, periodic term, and noise term. The first-order difference sign rate of change for each component is calculated. ,when If the value is less than a preset trend term discrimination threshold, the component is determined to be a trend term; power spectral density analysis is performed on the remaining components, and if there is a significant peak (the peak value exceeds a specified multiple of the average power, such as 3 times), it is determined to be a periodic term; the remaining components are classified as noise terms. In this embodiment, the tail portion is taken. Each component, or the component whose eigenvalue is less than 1% of the largest eigenvalue, is treated as a noise term.

[0029] Let the set of retained component indices be... Reconstruct the trajectory matrix:

[0030] By diagonal average Reducing the sequence to a one-dimensional value yields the upstream denoised sequence. .

[0031] S1.3, the improved CC method is used to jointly calculate the embedding dimension and delay time of the denoised sequence, specifically as follows: For a given embedding dimension and delay time The phase points in the reconstructed phase space are:

[0032] in, Let i be the i-th phase point, with dimension m. For phase point index, Let i be the value of the upstream denoised sequence at time point i. For the upstream denoised sequence in the first... The value at each point in time. For the upstream denoised sequence in the first... The value at each point in time. This represents the number of phase points.

[0033] The relevant integral is defined as:

[0034] in, The relevant integral takes values ​​in the range [0,1]. As the normalization factor, The length of the upstream denoised sequence. Let be the Euclidean distance between the i-th and j-th phase points. This is the Heaviside step function, and the neighborhood radius r takes multiple values, usually a multiple of the sequence standard deviation.

[0035] Define the test statistic:

[0036] Furthermore, the Delta statistic is defined as follows:

[0037] in, This represents the maximum value of the test statistic S among all neighborhood radius values. This represents the minimum value of the test statistic S among all neighborhood radius values.

[0038] Construct the objective function:

[0039] in, This represents the number of possible values ​​for the neighborhood radius.

[0040] Find the optimal parameter pair using grid search:

[0041] in, For the optimal embedding dimension, For the optimal delay time, This represents the minimum value parameter operator. The upper limit of the search for the embedding dimension. The upper limit of the search time is the delay time. To define the search interval for the embedding dimension, The search interval is the time delay.

[0042] S1.4, further, based on the determined optimal embedding dimension and optimal delay time Denoising the upstream sequence Phase space reconstruction is performed to obtain the upstream phase space trajectory. :

[0043] in, The number of phase points in the upstream phase space trajectory. Indicates that the upstream denoising sequence is at the 1st... The value at each time point.

[0044] S1.5, Electricity consumption sequence of downstream industries Repeat steps S1.1 to S1.4 to obtain the downstream phase space trajectory. .

[0045] S2, establish the differential homeomorphic mapping relationship between the upstream and downstream phase space trajectories, and construct a joint phase space based on the mapping relationship; S2.1, in this embodiment, establishing the differential homeomorphic mapping relationship between the upstream and downstream phase space trajectories specifically involves: S2.1.1, calculate the local tangent space coordinates of the upstream and downstream phase space trajectories. For the upstream phase space trajectory... Each phase point in The local tangent space coordinates are calculated using local principal component analysis. Specifically, phase points are selected. Find the k nearest neighbors, calculate the covariance matrix of these points, perform eigenvalue decomposition on the covariance matrix, take the first d principal component directions as the basis of the tangent space, and then... Projecting onto the tangent space yields the local tangent space coordinates. .

[0046] Similarly, for the downstream phase space trajectory Each phase point in The local tangent space coordinates were calculated using the same method. .

[0047] S2.1.2, further, based on the local tangent space coordinates, the Riemann curvature tensor of the upstream and downstream phase space trajectories is calculated. The Riemann curvature tensor is used to describe the degree of curvature of the manifold locally, serving as a geometric constraint for subsequent mappings.

[0048] For the upstream phase space trajectory, at the phase point At this point, the Riemann curvature tensor is calculated using the geometric relationship between the local tangent space coordinates and neighboring points, employing a curvature estimation method on the discrete manifold. Specifically, the surface is fitted using the second derivative information in the local tangent space, and the curvature tensor is extracted from the second fundamental form of the fitted surface. The Riemann curvature tensor is calculated using the same method for the downstream phase space trajectory. .

[0049] S2.1.3, calculate the mutual information entropy between all phase point pairs in the upstream and downstream phase space trajectories. For the phase points of the upstream phase space trajectory... The components of each dimension are discretized, and their value range is divided into B equally spaced intervals to obtain the discretized values. The phase point of the downstream phase space trajectory. Using the same discretization method, the values ​​are obtained. .

[0050] The mutual information entropy is defined as:

[0051] in, For phase point The values ​​after discretization For phase point The values ​​after discretization for and The joint probability distribution, for Marginal probability distribution, for The marginal probability distribution.

[0052] Calculate the mutual information entropy between all phase point pairs, and mark the top k phase point pairs with the largest entropy values ​​as the initial corresponding anchor point set. .

[0053] S2.1.4, using local tangent space coordinates as conformal invariants, the initial corresponding anchor point set as a hard constraint, and minimizing the difference in the Riemann curvature tensor as a curvature consistency constraint, constructs a conformal mapping distortion energy function. :

[0054] in, Let be the differential homeomorphic mapping function to be solved. For the conformal energy term, the local tangent space coordinates are taken as conformal invariants. For anchor point constraints, the initial corresponding anchor point set is used as a hard constraint. The curvature consistency term is constrained by minimizing the difference in the Riemann curvature tensor. These are the weighting coefficients for each energy term.

[0055] Furthermore, the gradient descent algorithm on the Riemannian manifold is used to iteratively solve for the energy function that distorts the conformal mapping. Minimized differential homeomorphism mapping function An initial mapping function is constructed using the initial corresponding anchor point set as a hard constraint, along the energy function. The mapping function is updated in the descent direction. The process continues until the change in the conformal mapping distortion energy function is less than a preset threshold or the maximum number of iterations is reached, at which point convergence is obtained, yielding the differential homeomorphic mapping function. .

[0056] S2.2, the formation of the joint phase space specifically refers to: Through the differential homeomorphism mapping function Upstream phase space trajectory Each phase point in the upstream phase space is mapped to the coordinate space of the downstream phase space trajectory to obtain the upstream projected trajectory. Upstream projection trajectory Each point in the trajectory of the downstream phase space The points in the array have the same spatial dimension.

[0057] Furthermore, the local reachability density ratio between the upstream projected trajectory and the downstream phase space trajectory at each time step is calculated, whereby the local reachability density is defined as the ratio between the point and its downstream phase space trajectory. The reciprocal of the reachable distance between the nearest neighbors:

[0058] in, For point Locally achievable density, This refers to the number of nearest neighbors in the calculation of local reachability density. For point of The set of nearest neighbors, For point Time The reachable distance.

[0059] Furthermore, the local reachability density ratio for each time index t is calculated. :

[0060] in, Upstream projection trajectory points Locally achievable density, Downstream phase space trajectory points The locally achievable density.

[0061] When the local reachability density ratio exceeds a preset threshold, the mapping point at that moment is identified as an abnormal mapping point. Abnormal mapping points are removed, and the remaining points of the upstream projection trajectory are mapped to the points of the downstream phase space trajectory according to their time indices, embedding them together into the same coordinate space to form a joint phase space. .

[0062] S3, perform continuous homology analysis on the joint phase space, and identify the filter parameter values ​​where homology characteristics change abruptly as time-varying boundary points; This step uses continuous cohomology analysis of the joint phase space to identify the moment when the industry chain linkage pattern undergoes a qualitative change from the perspective of topological structure evolution. Continuous cohomology is a topological data analysis method that captures the birth and death of topological features (such as connected components, loops, and holes) by observing the connectivity of data points at different scales. When the topological features change significantly, the corresponding filter parameter value is the time-varying boundary point.

[0063] In this embodiment, the step of performing continuous homology analysis on the joint phase space and identifying the filter parameter value where the homology feature undergoes abrupt change as the time-varying boundary point specifically involves: S3.1 It should be noted that the weighted simplicium is a topological description of the connections between data points. Its construction process is as follows: each trajectory point is considered a vertex. When the distance between two vertices is less than a given threshold, an edge connects the two vertices; when the distance between any two pairs of three vertices is less than the threshold, a triangle is filled; and so on, forming a high-dimensional structure. These vertices, edges, triangles, and their high-dimensional generalizations are collectively referred to as simplicium. Weighting refers to incorporating tangent space angle information into the distance function, allowing the connections to consider both spatial location and local geometric direction, thereby improving sensitivity to dynamic pattern changes.

[0064] For the joint phase space Any two points in and Construct a weighted distance function as the filtering parameter:

[0065] in, The weighted distance is used as a filtering parameter. These are the weighting coefficients. for and The Euclidean distance between them and Points and The principal direction of the tangent space at a given location is obtained through local principal component analysis (taking the first principal component in the neighborhood). The dot product of vectors. and Let L2 norm be the vector. As a normalization factor, the included angle is normalized to [0,1]. For the inverse cosine function, convert the inner product value to radians.

[0066] Furthermore, using weighted distance As filtering parameters, generate Vietoris-Rips simple complex filtering sequences. For each filter parameter value Rips complex Defined as: Using all points on the trajectory as vertices (0-dimensional simplex), when two points... , satisfy When the two points are connected by an edge (1-dimensional simplex); when any three points are paired, the following condition is met: When filling a simplex, a triangle is filled (2-dimensional simplex); the same applies to higher-dimensional simplexes (tetrahedrons, etc.).

[0067] With filter parameters Gradually increase from 0 to Simple complexes gradually grow from discrete point sets to form structures such as edges and triangles, thus forming a filtering sequence.

[0068] It should be noted that by introducing the tangent space angle component, the weighting function ensures that: only when two points are spatially adjacent and have the same local motion direction will they be connected with a smaller filtering parameter; conversely, if two points are spatially adjacent but have orthogonal motion directions, a larger filtering parameter is required to connect them. This characteristic makes subsequent continuous cohomology analysis more sensitive to abrupt changes in dynamic patterns, thereby improving the accuracy of time-varying boundary point identification.

[0069] S3.2, the step of calculating persistent homology groups on the filtered sequence to obtain the persistent sequences of each homology class specifically involves: on the filtered sequence... The persistence coherence groups of each dimension are calculated, the birth time and death time of each coherence class are recorded, and a persistence graph is constructed. The horizontal axis represents the birth time, the vertical axis represents the death time, the points near the diagonal represent features that appear briefly (usually noise), and the points far from the diagonal represent features that are persistent.

[0070] Furthermore, the persistence of each homology class is calculated, which is the difference between the time of death and the time of birth. The persistence is then sorted in ascending order by the time of birth to obtain the persistence sequence.

[0071] Furthermore, change point detection is performed on persistent sequences to identify locations where statistically significant mutations occur. Change point detection can employ the Pettitt test, binary segmentation, or Bayesian change point detection. For each detected change point location, the birth time of the corresponding homology class is taken as the time-varying boundary point.

[0072] S4. The joint phase space is divided into multiple sub-regions according to the time-varying boundary point. The optimal transmission distance is calculated in each sub-region to obtain the dynamic time delay parameter function. In this embodiment, the process of dividing the joint phase space into multiple sub-regions based on the time-varying boundary point, calculating the optimal transmission distance in each sub-region, and obtaining the dynamic time delay parameter function specifically involves: S4.1, let the set of time-varying boundary points obtained by identification be... ,in, The number of time-varying boundary points. The division into continuous sub-regions using time-varying boundary points as dividing boundaries specifically involves: using the time-varying boundary points as dividing boundaries, the upstream trajectory point sequence (by time index) in the joint phase space is divided into... A contiguous sub-region.

[0073] For sub-regions Suppose that the sub-region contains upstream trajectory points and There are downstream trajectory points. Since upstream and downstream trajectory points correspond according to time index, in the sub-region Within, both the upstream timestamp set and the downstream timestamp set are... The time index value in the data.

[0074] Define the timestamp distribution of the upstream trajectory points as the source distribution. The timestamp distribution of downstream trajectory points is defined as the target distribution. In this embodiment, both the source distribution and the target distribution are uniform, meaning that each timestamp is assigned the same probability mass.

[0075] Furthermore, the transmission cost matrix is ​​defined. , of which elements This indicates that the upstream timestamp Transmitted to downstream timestamps The cost, in this embodiment, is the square of the time offset:

[0076] Construct an entropy-regularized optimal transport problem and solve for the probability transition matrix. :

[0077] in, This represents the optimal probability transition matrix. This indicates that the upstream timestamp Transmitted to downstream timestamps The probability, Let represent the set of all probability transition matrices that satisfy the marginal distribution constraints. is the entropy regularization coefficient, which controls the smoothness of the solution. is the entropy function, used to encourage the smoothness of the probability transition matrix.

[0078] Furthermore, the Sinkhorn algorithm is used iteratively to solve the above entropy-regularized optimal transport problem. The Sinkhorn algorithm converges quickly to the optimal solution by alternating row and column normalization, yielding the optimal probability transition matrix. The iterative solution of the Sinkhorn algorithm adopts a mature implementation method in existing technology, which will not be elaborated here.

[0079] S4.2, for sub-regions From the optimal probability transition matrix The total transmission probability corresponding to each possible time offset is calculated, and the time offset with the highest total transmission probability is taken as the local time delay constant of that sub-region. :

[0080] in, Time offset, local time delay constant This represents the average time delay by which changes from upstream to downstream are transmitted within this sub-region.

[0081] S4.3, using the center time point of each sub-region (the average of the start and end times of that sub-region) as interpolation nodes, cubic spline interpolation is performed on the local time delay constants of each sub-region. Specifically, a cubic spline interpolation function is constructed with the center time point of each sub-region as the x-axis and the corresponding local time delay constant as the y-axis. Cubic spline interpolation ensures that the interpolation function is continuous at the nodes and has continuous first and second derivatives, thus obtaining a smooth time delay variation curve. The interpolated function... This is a dynamic time-delay parameter function that changes continuously with time, where t is the time index.

[0082] S5 embeds the dynamic time delay parameter function into the time-series neural network, jointly optimizes the network weights, and outputs the predicted electricity consumption value of downstream industries.

[0083] In this embodiment, the step of embedding the dynamic time delay parameter function into a time-series neural network, jointly optimizing the network weights, and outputting the predicted electricity consumption value for downstream industries specifically involves: The obtained dynamic time delay parameter function As a differentiable time-shifting operator, it is embedded in the input layer of a temporal neural network. This operator enables the network to perform time shifting at each time step. ,according to The value, from the upstream historical sequence Dynamically extract aligned feature vectors Specifically, the aligned feature vectors are obtained by analyzing the upstream sequence at time points. Linear interpolation is performed at the point to obtain: like Integer ,but ; like ,in ,but ; like ,but ; like ,but .

[0084] because Continuously differentiable, this linear interpolation operator pairs The partial derivatives of the time delay parameter exist and can participate in backpropagation, enabling joint optimization of the time delay parameter and network weights.

[0085] Aligned feature vectors By combining it with historical electricity consumption data from downstream sources, a joint input feature is formed. .Will Input the core layer of the temporal neural network. This implementation uses a gated recurrent unit (GRU) network, configured as a stack of two layers, with 64 hidden units in each layer and a dropout rate of 0.2. The forward propagation calculation is as follows:

[0086]

[0087] in, and These are the hidden states of the first and second layers at the previous time step (the initial state is a zero vector).

[0088] Furthermore, the output of the second-layer GRU... Inputting the fully connected layer yields the preliminary prediction for the current time step:

[0089] in, This is the output layer weight matrix. This is the output layer bias.

[0090] Constructing a joint loss function End-to-end training is performed, and the function includes a prediction error term and a time delay function smoothness constraint term:

[0091] in, The number of training samples, For time indexing, This is the time index set of the training set. This represents the actual electricity consumption of downstream industries. The regularization coefficient controls the strength of the smoothness constraint. The sequence length is given.

[0092] Backpropagation is performed using the Adam optimizer to calculate the loss function with respect to network weights and time delay parameters. The gradients are calculated, and these parameters are updated. The optimizer configuration is shown in Table 1.

[0093] Table 1

[0094] Iterate through training until the loss function converges. The convergence condition is: the validation set loss does not decrease for 10 consecutive rounds, and the change in training set loss is less than 1 / 2. Or reach the maximum number of training rounds (e.g., 500).

[0095] After training, the latest upstream industry electricity consumption observation data is input into the trained time-series neural network. The network dynamically extracts aligned feature vectors from the latest upstream electricity consumption data based on the dynamic time lag value corresponding to the current prediction time, and then combines this with the latest downstream industry data. The system uses historical electricity consumption data from each step, and through forward propagation calculations in the core layer and fully connected layer, it ultimately outputs the predicted electricity consumption value for the downstream industry at the next moment (or multiple subsequent moments). For multi-step prediction, an autoregressive approach is adopted: for each prediction step, the predicted value is used as the downstream historical input for subsequent predictions.

[0096] Example 2, Figure 2 The present invention provides a system for predicting industry electricity consumption based on industrial chain linkages, comprising: The data processing unit is used to acquire the electricity consumption sequences of upstream and downstream industries, and construct the phase space trajectories of upstream and downstream industries respectively through delayed embedding. A joint phase space construction unit is used to establish a differential homeomorphic mapping relationship between upstream and downstream phase space trajectories, and to construct a joint phase space based on the mapping relationship. The boundary point identification unit is used to perform continuous homology analysis on the joint phase space and identify the filter parameter values ​​where the homology characteristics change abruptly as time-varying boundary points. The time delay calculation unit is used to divide the joint phase space into multiple sub-regions according to the time-varying boundary point, calculate the optimal transmission distance in each sub-region, and obtain the dynamic time delay parameter function; The prediction output unit is used to embed dynamic time-delay parameter functions into a time-series neural network, jointly optimize the network weights, and output the predicted electricity consumption value for downstream industries.

[0097] Example 3, an electronic device, such as Figure 3 As shown, it includes a memory and a processor: the memory is used to store a program; the processor is used to execute the program to implement any of the embodiments in Example 1.

[0098] Since the electronic device described in this embodiment is the device used to implement the method in Embodiment 1 of the present invention, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in Embodiment 1 of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any device used by those skilled in the art to implement the method in the embodiments of this application falls within the scope of protection of this application.

[0099] Example 4: A readable storage medium having a computer program stored thereon, such as... Figure 4 As shown, when the computer program is executed by the processor, it implements any of the embodiments in Example 1.

[0100] Example 5: To further verify the predictive effectiveness of this method, this example uses monthly electricity consumption data from a certain province from January 2022 to December 2023, a total of 24 months. The predicted values ​​of this method are compared with the actual values, and also compared with the prediction results of the Transformer method.

[0101] like Figure 5 As shown in the figure, three curves are displayed. It can be seen that the prediction curve of the method of this invention closely follows the actual value curve throughout the entire time period, while the Transformer prediction curve deviates significantly in multiple time periods. Especially during the period of large fluctuations in electricity consumption in the second half of 2022, the prediction trajectory of this method remained highly consistent with the actual value.

[0102] Peak and valley fluctuation capture capability: During the peak electricity consumption period in July and August 2022 (the actual values ​​reached 14.23 billion kWh and 14.52 billion kWh), the predicted values ​​of this method were 14.15 billion kWh and 14.48 billion kWh, respectively, with the error controlled within 1%; the predicted values ​​of Transformer were 13.89 billion kWh and 14.25 billion kWh, respectively, with errors of 2.4% and 1.9%.

[0103] Sudden change response capability: During the sharp drop in electricity consumption in December 2022 (from 12.03 billion kWh to 10.58 billion kWh) and the continued decline in January 2023 (to 9.56 billion kWh), this method captured this trend in a timely manner, with predicted values ​​of 10.62 billion kWh and 9.61 billion kWh, respectively; while the predicted values ​​of Transformer (10.32 billion kWh and 9.28 billion kWh) had a lag bias of about 2-3 percentage points.

[0104] Seasonal fluctuation restoration: During the peak electricity consumption period in the summer of 2023 (July-August), the actual values ​​were 13.89 billion kWh and 14.52 billion kWh, respectively. The predicted values ​​by this method were 13.95 billion kWh and 14.48 billion kWh, respectively, with relative errors of 0.43% and 0.28%. The predicted values ​​by Transformer were 13.68 billion kWh and 14.25 billion kWh, respectively, with relative errors of 1.51% and 1.86%.

[0105] To further quantify and compare the prediction performance of the two methods, four indicators were used for evaluation: mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The quantitative comparison results on the test set are shown in Table 2.

[0106] Table 2

[0107] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0108] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0109] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0110] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0111] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0112] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for forecasting industry electricity consumption based on supply chain linkages, characterized in that, Includes the following steps: Obtain the electricity consumption sequences of upstream and downstream industries, and construct the phase space trajectories of upstream and downstream industries respectively through delayed embedding; Establish a differential homeomorphic mapping relationship between upstream and downstream phase space trajectories, and construct a joint phase space based on the mapping relationship; Continuous homology analysis is performed on the joint phase space, and the filter parameter values ​​where the homology characteristics change abruptly are identified as time-varying boundary points; The joint phase space is divided into multiple sub-regions based on the time-varying boundary point. The optimal transmission distance is calculated in each sub-region to obtain the dynamic time delay parameter function. By embedding dynamic time delay parameter functions into a time-series neural network and jointly optimizing network weights, the predicted electricity consumption values ​​for downstream industries are output.

2. The industry electricity consumption forecasting method based on industrial chain linkages according to claim 1, characterized in that, The construction of upstream and downstream phase space trajectories specifically involves: The electricity consumption series of upstream industries is denoised to obtain the upstream denoised series; An improved CC method is used to jointly calculate the embedding dimension and delay time for the denoised sequence; Based on the embedding dimension and delay time, the upstream denoised sequence is reconstructed in phase space to obtain the upstream phase space trajectory. Repeat the above steps for the electricity consumption sequence of downstream industries to obtain the downstream phase space trajectory.

3. The industry electricity consumption forecasting method based on industrial chain linkages according to claim 2, characterized in that, The establishment of the differential homeomorphism mapping relationship specifically includes: Calculate the local tangent space coordinates and Riemann curvature tensor of each phase point in the upstream and downstream phase space trajectories; Calculate the mutual information entropy between all phase point pairs in the upstream and downstream phase space trajectories, and mark the k phase point pairs with the largest entropy values ​​as the initial corresponding anchor point set; Using local tangent space coordinates as conformal invariants, the initial corresponding anchor point set as a hard constraint, and minimizing the difference of the Riemann curvature tensor as a curvature consistency constraint, a conformal mapping distortion energy function is constructed. The differential homeomorphism that minimizes the energy function is solved by gradient descent on the Riemannian manifold.

4. The industry electricity consumption forecasting method based on industrial chain linkage as described in claim 3, characterized in that, The formation of the joint phase space specifically refers to: The upstream phase space trajectory is projected onto the coordinate space of the downstream phase space trajectory through the differential homeomorphic mapping function, eliminating mapping points with abnormally high reachability density ratios. The remaining upstream projection trajectories and downstream phase space trajectories are matched by time index and embedded together into the same coordinate space to form a joint phase space.

5. The industry electricity consumption forecasting method based on industrial chain linkage as described in claim 4, characterized in that, The identification of the time-varying boundary point is specifically as follows: Transform the trajectory points in the joint phase space into a weighted simplex and generate a filtering sequence; Calculate persistent homology groups on the filtered sequences to obtain the persistent sequences of each homology class; Change point detection is performed on persistent sequences, and the filter parameter value where a mutation occurs in the persistence is used as the time-varying boundary point.

6. The industry electricity consumption forecasting method based on industrial chain linkage as described in claim 5, characterized in that, The obtained dynamic time delay parameter function specifically includes: The region is divided into continuous sub-regions using the time-varying boundary point as the dividing line. Within each sub-region, the probability transition matrix of the timestamps of the upstream and downstream trajectory points is obtained by solving the optimal transmission with entropy regularization, and the time offset with the highest transmission probability is extracted as the local time delay constant. The local time delay constants of adjacent sub-regions are smoothly connected by cubic spline interpolation to obtain a dynamic time delay parameter function that changes continuously with time.

7. The industry electricity consumption forecasting method based on industrial chain linkage as described in claim 6, characterized in that, The output of downstream industry electricity consumption forecasts specifically includes: The dynamic time delay parameter function is embedded as the input layer of a time-series neural network as a differentiable time offset operator, and the alignment features are extracted from the upstream historical sequence using the operator. The alignment features are concatenated with downstream historical electricity consumption data and then input into the network to obtain preliminary predicted values. A joint loss function is constructed for end-to-end training. After iterative training until the loss function converges, the predicted electricity consumption value of downstream industries is calculated using the latest upstream electricity consumption data as input.

8. A system using the industry electricity consumption forecasting method based on industrial chain linkage as described in any one of claims 1-7, characterized in that, include: The data processing unit is used to acquire the electricity consumption sequences of upstream and downstream industries, and construct the phase space trajectories of upstream and downstream industries respectively through delayed embedding; A joint phase space construction unit is used to establish a differential homeomorphic mapping relationship between upstream and downstream phase space trajectories, and to construct a joint phase space based on the mapping relationship. The boundary point identification unit is used to perform continuous homology analysis on the joint phase space and identify the filter parameter values ​​where the homology characteristics change abruptly as time-varying boundary points. The time delay calculation unit is used to divide the joint phase space into multiple sub-regions according to the time-varying boundary point, calculate the optimal transmission distance in each sub-region, and obtain the dynamic time delay parameter function; The prediction output unit is used to embed dynamic time delay parameter functions into a time-series neural network, jointly optimize network weights, and output the predicted electricity consumption value of downstream industries.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor: The memory is used to store programs; The processor is used to execute the program to implement each step of the industry electricity consumption forecasting method based on industrial chain linkage as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the industry electricity consumption forecasting method based on industrial chain linkage as described in any one of claims 1 to 7.

Images