A method for simulating energy consumption behavior based on a generative model

Abstract

The application discloses a kind of energy consumption behavior simulation methods based on generating model, comprising the following steps: S1, by the space-time energy consumption data of equipment acquisition, data interpolation reconstruction is carried out and is converted into high-dimensional space-time data matrix;S2, data matrix is decomposed into multiple scale sub-matrix, using dynamic noise driven forward diffusion process obtains space-time scale coding data;S3, using cross-scale data overlap fusion and dynamic noise weight decreasing mechanism, execute reverse diffusion process, reconstruct space-time energy consumption initial data;S4, according to reconstruction data, construct characteristic space attraction field and determine the spatial position of quantum chaos vultur group;S5, dynamically adjust group position, make it follow attraction field gradient and carry out adaptive optimization, obtain dynamic optimization sequence.The application is combined by quantum chaos optimization and space-time diffusion model, can accurately simulate complex and dynamic energy consumption behavior, has strong optimization effect and wide application prospect.

Description

Technical Field

[0001] This invention relates to the field of energy consumption calculation technology, and in particular to a method for simulating energy consumption behavior based on a generative model. Background Technology

[0002] With the rapid development of the information society, energy consumption monitoring and management has become a critical issue that various industries urgently need to address. Particularly in the fields of smart grids, smart buildings, and industrial automation, accurately simulating and predicting the energy consumption behavior of equipment and users has become a key factor in improving energy efficiency, reducing waste, and optimizing resource allocation. Currently, energy consumption behavior simulation technology is gradually developing, but it still faces many challenges.

[0003] Currently, research on energy consumption behavior simulation mainly focuses on technical approaches based on traditional statistical methods, machine learning algorithms, and system dynamics models. For example, energy consumption models based on statistical analysis methods typically predict future energy consumption patterns using historical data, but their accuracy is limited because they cannot fully account for the complex dynamic changes in equipment and user behavior. Machine learning algorithms, especially deep learning techniques, have been applied to energy consumption behavior prediction, building models by learning from large amounts of historical energy consumption data. However, these methods often suffer from insufficient ability to model complex systems and nonlinear behaviors, and their generalization ability is also limited. System dynamics models, by establishing equations describing the interactions of the system's components, provide relatively accurate simulation results, but this method has high computational complexity, requires a large amount of historical data, and often performs poorly when dealing with non-stationarity in complex environments.

[0004] Furthermore, existing energy consumption behavior simulation techniques often struggle to adapt to complex and dynamically changing energy consumption characteristics when dealing with real-time dynamic changes in users or equipment. Most methods fail to adequately consider the impact of spatiotemporal factors on energy consumption behavior during the simulation process, resulting in limitations in the accuracy and adaptability of the models when faced with complex, nonlinear, and variable energy consumption behaviors.

[0005] In recent years, energy consumption behavior simulation based on spatiotemporal data has gradually attracted attention, and the introduction of spatiotemporal diffusion models and quantum algorithms has provided new ideas. Spatiotemporal diffusion models can handle multidimensional data and further improve the accuracy of energy consumption behavior prediction through joint modeling of spatial and temporal dimensions. However, existing spatiotemporal diffusion models mostly focus on static data analysis, and there is still considerable room for improvement in dynamic real-time data processing, model optimization, and multi-scale fusion.

[0006] Quantum algorithms, especially quantum chaotic algorithms, have made some progress in applications across multiple fields. However, their unique advantages in solving optimization and search problems of complex systems have not yet been fully realized. Although some studies have attempted to apply quantum chaotic algorithms to optimization models, their combined application with spatiotemporal diffusion models, particularly for the accurate simulation and prediction of energy consumption behavior, has not yet become a mainstream research direction.

[0007] Therefore, how to provide a method for simulating energy consumption behavior based on generative models is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose an energy consumption behavior simulation method based on a generative model. This invention addresses the shortcomings of existing energy consumption behavior simulation technologies in real-time dynamic optimization, multi-scale data processing, and spatiotemporal feature modeling, and has higher accuracy and application value.

[0009] An energy consumption behavior simulation method based on a generative model according to an embodiment of the present invention includes the following steps:

[0010] S1. Based on the overall continuity characteristics of the spatiotemporal energy consumption data of the collected equipment, the locations of data anomalies and missing data are dynamically determined point by point, and bidirectional interpolation reconstruction is performed. A high-dimensional continuous spatiotemporal data matrix is ​​generated through dynamic data scale transformation.

[0011] S2. Based on a high-dimensional continuous spatiotemporal data matrix, the data matrix is ​​alternately decomposed into multi-scale sub-matrices along the spatiotemporal axis. The forward diffusion process of each scale sub-matrix is ​​driven by random dynamic noise to obtain spatiotemporal scale multi-dimensional encoded data.

[0012] S3. Based on multi-dimensional encoded data, the reverse diffusion process is performed by using cross-scale data overlap fusion and dynamic noise weight reduction mechanism to realize the layer-by-layer restoration of encoded data at each scale under multi-scale fusion constraints, and obtain the reconstructed spatiotemporal energy consumption initial data.

[0013] S4. Based on the reconstructed initial spatiotemporal energy consumption data, construct the characteristic spatial gravitational field of the spatiotemporal energy consumption data, and determine the spatial location distribution of the initial quantum chaotic African vulture group.

[0014] S5. Within the gravitational field of the characteristic space, the positions of individual members of the quantum African vulture group are updated in a dynamic chaotic probability manner, so that the group positions are dynamically adjusted to adapt to the gravitational gradient of the spatiotemporal data, thereby obtaining a dynamic optimization sequence of model parameters.

[0015] S6. Utilizing the dynamic optimization sequence, the cross-scale data overlap fusion and dynamic noise weight reduction mechanism are adjusted in reverse to perform closed-loop adaptive optimization of the multi-scale reverse diffusion reconstruction process, thereby obtaining a dynamically optimized spatiotemporal energy consumption behavior characteristic sequence.

[0016] Optionally, S1 specifically includes:

[0017] S11. Based on the raw spatiotemporal energy consumption data collected by the equipment, arrange the data in the continuous order of the data collection time to form an initial data sequence, calculate the local density value between each data point and the data points at multiple times in its neighborhood, and dynamically determine the location of data anomalies in the initial data sequence based on the gradient change of the local density value over time.

[0018] S12. Based on the location of outliers, the initial data sequence is divided into several continuous data segments with adjacent outlier locations as the dividing boundaries. The integrity of each data segment is evaluated separately. Based on the sampling time interval of the data points in the time dimension, the specific location of the missing data points within the segment is dynamically determined.

[0019] S13. Using the missing data locations as interpolation targets, establish a spatiotemporal neighborhood window centered on the missing data points, and calculate the information entropy values ​​of adjacent data points within each neighborhood window. By gradually minimizing the information entropy differences among data points within the neighborhood, determine the interpolation weights of the interpolation target points in both spatiotemporal dimensions, perform bidirectional adaptive interpolation reconstruction, and generate reconstructed data values ​​for each missing location.

[0020] S14. After repairing the missing positions of each data sub-segment with reconstructed data values, a temporally continuous sub-segment overlap area is formed by overlapping sub-segment boundaries. Data smoothing weights are established through the numerical differences between data points in the boundary overlap area. Adaptive numerical smoothing fusion of the boundary area is performed on each sub-segment to keep the data values ​​between adjacent sub-segments continuous and form a continuous spatiotemporal data sequence.

[0021] S15. Using continuous spatiotemporal data sequences, a two-dimensional spatiotemporal data matrix is ​​constructed by mapping the data points in the data sequence to spatiotemporal two-dimensional coordinates. The numerical distribution imbalance of the data points in the data matrix is ​​evaluated in a local region. The dynamic conversion weight coefficient of the data scale is determined based on the ratio of the numerical variance in the local region to the mean of the neighborhood. The numerical scale of the data points in the data matrix is ​​dynamically adjusted region by region.

[0022] S16. Based on the scaled data matrix, perform high-dimensional nonlinear spatial expansion mapping of the data row by row and column by column. Using the row and column positions of each data point in the original matrix as the mapping basis, generate multiple high-dimensional subspace mapping data points through nonlinear mapping functions, and recombine the mapping data points to form a high-dimensional continuous spatiotemporal data matrix.

[0023] Optionally, S2 specifically includes:

[0024] S21. Based on the spatiotemporal coordinates of each data point in the high-dimensional continuous spatiotemporal data matrix, a progressive non-uniform density threshold segmentation strategy is adopted along the time axis and the spatial axis to calculate the spatiotemporal data density difference in the local region of the matrix. Based on the gradient change of density difference, the matrix is ​​dynamically decomposed into multiple data sub-matrices with progressively changing scale and irregular boundaries.

[0025] S22. Based on each data submatrix, calculate the spatial correlation coefficient of each data point within the submatrix relative to its local spatiotemporal neighbor data points, and determine the spatial clustering center position of the data within the submatrix based on the spatial correlation coefficient, and generate submatrix spatial clustering maps that reflect the spatial distribution structure of the data one by one.

[0026] S23. Based on the cluster center positions of each submatrix spatial clustering mapping and the spatial correlation coefficients of each data point, the normalization scaling factor is determined point by point. The normalization scaling factor is determined by the nonlinear function of the spatial correlation coefficients between the data points and the cluster center positions. The submatrix is ​​spatially adaptively scaled and normalized in a point-by-point adjustment manner to obtain the scaled normalized data submatrix.

[0027] S24. Utilize the data density distribution within the data submatrix after scale normalization, dynamically determine the spatial distribution of random noise variance by the change in the spatial density gradient of the local neighborhood where each data point is located, and generate a dynamic random noise data matrix with adaptive variance based on the local data density gradient, so that the dynamic random noise within each submatrix exhibits a non-uniform distribution in the spatial dimension.

[0028] S25. The dynamic random noise data matrix is ​​superimposed one by one on the corresponding scale-normalized data submatrix. Based on the superimposed data submatrix, the local spatial density difference of the data points is calculated point by point. The diffusion intensity factor in the forward diffusion process is determined according to the spatial density difference. The diffusion degree of each data point is gradually adjusted, and the stepwise forward diffusion process between data points is executed.

[0029] S26. During the forward diffusion process, based on the diffusion state change process of data points in the data submatrix at each scale, the data density change trajectory of each data point in the local spatiotemporal neighborhood during the diffusion process is sampled and recorded in real time. Based on the data point density change trajectory, a multidimensional data encoding sequence containing spatiotemporal scale and diffusion state is generated to obtain spatiotemporal scale multidimensional encoded data.

[0030] Optionally, S3 specifically includes:

[0031] S31. Based on the spatiotemporal scale multidimensional encoded data matrix, calculate the spatial position offset of all data points in each scale encoded data matrix relative to the global data matrix, and determine the scale correlation strength of each data point based on the spatial position offset.

[0032] S32. Determine the overlapping area of ​​cross-scale data based on the scale correlation strength, and calculate the numerical coherence degree of corresponding data points between different scale coded data matrices in the overlapping area one by one. The numerical coherence degree is a nonlinear coherence function of the numerical difference of data points.

[0033] S33. Construct a data fusion path point by point based on the numerical synergy degree, taking the data point with the highest numerical synergy degree as the starting point of the fusion path, and gradually extending to the data points with lower numerical synergy degree, dynamically constructing a data fusion tree topology structure.

[0034] S34. Utilize the data fusion tree topology structure to perform cross-scale numerical fusion of data nodes in the tree topology structure one by one, so as to realize the step-by-step data fusion of data matrices of different scales under the fusion topology structure and form a cross-scale fusion data matrix.

[0035] S35. Based on the cross-scale fusion data matrix, calculate the local stability index of each data point and its spatial neighborhood, and dynamically generate a random noise weight matrix with non-uniform spatial distribution based on the gradient of the local stability index.

[0036] S36. The random noise weight matrix is ​​superimposed point by point onto the cross-scale fused data matrix. Based on the local stability index gradient of each data point in the superimposed matrix, the adaptive deceleration rate of the back diffusion intensity is determined point by point to dynamically adjust the gradual decrease process of the random noise weight.

[0037] S37. Based on the gradual decrease of random noise weight, reverse diffusion is performed point by point. The data fusion process is traced back layer by layer based on the fused data topology path. The initial values ​​before data diffusion are restored point by point, and finally the reconstructed spatiotemporal energy consumption initial data is obtained.

[0038] Optionally, S33 specifically includes:

[0039] S331. For each data point in the cross-scale data overlap area, take the data point as the reference point and calculate the spatial topological distance between the reference point and all other data points in the area. The spatial topological distance is defined as a nonlinear weighted combination function of the Euclidean distance of spatiotemporal coordinates and the difference in numerical coherence.

[0040] S332. Using the calculated spatial topological distance, determine the fusion topological centrality index of each data point in the cross-scale data overlap region one by one. The fusion topological centrality index is obtained by non-linear weighted summation of the reciprocal of the spatial topological distance between the data point and all other data points in the region.

[0041] S333. Based on the fusion topology centrality index, perform global sorting and dynamically determine the data point with the highest fusion topology centrality index as the starting node of the fusion topology structure.

[0042] S334. Taking the starting node of the fused topology as the origin, the topological diffusion potential energy from the starting node to its spatial neighborhood data points is calculated step by step. The topological diffusion potential energy is obtained by a nonlinear combination function of the fused topological centrality index and the spatial topological distance, and the data point with the largest topological diffusion potential energy is taken as the next extension node of the fused path.

[0043] S335. Repeat the calculation of topological diffusion potential energy for each determined extension node, and gradually determine the next extension node based on the real-time change of topological diffusion potential energy. Dynamically adjust the topological extension direction of the fusion path and gradually generate multiple branch structure data fusion paths in a nonlinear manner.

[0044] S336. The fusion order of each node in the real-time computing data fusion path structure is a nonlinear sorting function that combines the node's topological diffusion potential energy and topological centrality index. The hierarchical position of each node on the data fusion path is determined based on the fusion order.

[0045] S337. Based on the determined data fusion path and fusion order, connect the topological relationships between data nodes level by level, establish a data fusion tree network with asymmetric topological branch structure point by point, and obtain a refined, nonlinear and multi-branch structure of cross-scale data fusion tree topology.

[0046] Optionally, S35 specifically includes:

[0047] S351. Taking each data point in the cross-scale fused data matrix as the center, determine the spatial neighborhood window centered on the data point. The size of the neighborhood window is a fixed spatial radius range centered on the data point.

[0048] S352. Calculate the difference between the data point values ​​in the spatial neighborhood window and the center data point value one by one, and determine the degree of numerical deviation between the center data point and the spatial neighborhood data points based on the difference.

[0049] S353. Based on the degree of numerical deviation of all data points in the neighborhood window, determine the local stability index of the central data point point by point.

[0050] S354. Based on the local stability index of all data points in the cross-scale fused data matrix, calculate the stability index gradient of each data point in the spatial dimension. The stability index gradient is defined as the spatial weighted average of the difference between the local stability index of each data point and the local stability index of its neighboring data points.

[0051] S355. Based on the stability index gradient of the data points, determine the spatial distribution intensity of the random noise weights point by point.

[0052] S356. Based on the spatial distribution intensity of the calculated random noise weight, generate random noise values ​​point by point. The random noise values ​​follow a normal distribution with the random noise weight as the standard deviation and the mean as zero.

[0053] S357. Based on the cross-scale fusion data matrix, the generated random noise values ​​are superimposed point by point onto the corresponding data points to finally generate a random noise weight matrix with a non-uniform spatial distribution.

[0054] Optionally, S4 specifically includes:

[0055] S41. Based on the reconstructed initial spatiotemporal energy consumption data matrix, calculate the nonlinear combination of spatiotemporal distance and numerical difference between each data point and other data points in the matrix point by point, so as to determine the characteristic spatial potential energy value of each data point.

[0056] S42. Based on the characteristic spatial potential energy value of each data point, determine the potential energy spatial gradient vector of each position in the data matrix one by one through the spatial position coordinates. The magnitude and direction of the spatial gradient vector are clearly determined by the spatial partial derivative of the characteristic spatial potential energy value.

[0057] S43. Based on the potential energy spatial gradient vector at each position in the data matrix, calculate and determine the spatial derivative of the potential energy gradient vector point by point to generate the spatial rate of change matrix of the potential energy gradient vector.

[0058] S44. Based on the potential energy gradient vector space rate of change matrix, calculate the characteristic space gravitational intensity at each position in the data matrix space. The gravitational intensity value is determined by the absolute value of the value at each position in the potential energy gradient vector space rate of change matrix.

[0059] S45. Based on all the gravitational intensities of the characteristic space determined in the data matrix, construct a clear characteristic space gravitational field point by point. The field strength at each position in the characteristic space gravitational field is directly represented by the numerical value of the characteristic space gravitational intensity at the corresponding position.

[0060] S46. Based on the gravitational intensity distribution characteristics within the characteristic space gravitational field, and taking the relative magnitude of the local gravitational intensity as the basis, determine several spatial gravitational center nodes in the gravitational field. The gravitational center node is defined as the location point where the local gravitational intensity is extremely large in the characteristic space gravitational field.

[0061] S47. Based on the determined position of the spatial gravitational center node and the gravitational field strength, dynamically determine the spatial position distribution of the initial quantum chaotic African vulture group, and determine the specific spatial position coordinates of the initial quantum chaotic African vulture group by non-uniform random sampling according to the position distribution.

[0062] Optionally, S5 specifically includes:

[0063] S51. Based on the spatial position coordinates of each individual in the initial quantum chaotic African vulture group in the characteristic space gravitational field, calculate the spatiotemporal gravitational gradient vector between each vulture individual and all gravitational center nodes in the gravitational field, and form the global gravitational perception vector of the vulture individual.

[0064] S52. Based on the global gravitational perception vector of an individual vulture, the local chaotic perturbation probability at the current position is calculated in real time for each individual vulture. The local chaotic perturbation probability is determined by the nonlinear weighted sum of the change rates of the gravitational gradient vector at each position in the gravitational field of the characteristic space.

[0065] S53. Based on the local chaotic perturbation probability, dynamically determine the perturbation threshold of the position update direction for each vulture individual, and use the perturbation threshold to adjust the position update step size of the vulture individual along the gravitational gradient vector direction in real time.

[0066] S54. Calculate the dynamic spatial topological distance between the current spatial position of each vulture individual and the positions of each gravitational center node in the characteristic spatial gravitational field. The dynamic spatial topological distance considers a combination of Euclidean distance of position coordinates and nonlinear function of gravitational intensity difference.

[0067] S55. Based on the calculated dynamic spatial topological distance, the target gravitational node of each vulture individual is dynamically determined. The target gravitational node is adjusted in real time and defined as the gravitational center node with the smallest calculated value of the combination of topological distance and gravitational intensity.

[0068] S56. Each vulture individual is updated with its spatial coordinates gradually through a perturbation threshold, based on a dynamically determined target gravity node. The update method is constrained in real time by the combined effect of the gravity gradient vector and the probability of local chaotic perturbation.

[0069] S57. Record the updated spatial coordinates of each vulture individual in real time, form a trajectory sequence for dynamic adjustment of the group position one by one, and finally obtain the vulture group position distribution trajectory sequence that dynamically changes with the characteristic spatial gravitational field, and determine the position distribution trajectory sequence as the dynamic optimization sequence of model parameters.

[0070] Optionally, S55 specifically includes:

[0071] S551. For each vulture individual's current position, construct multiple spatiotemporal paths from the vulture individual to each gravitational center node in the characteristic spatial gravitational field, and calculate the cumulative value of the spatial gravitational gradient of all data points on each path. The cumulative value of the spatial gravitational gradient is the point-by-point summation of the absolute values ​​of the gravitational gradient at each data point on the path.

[0072] S552. Based on the cumulative value of the spatial gravitational gradient of each spatiotemporal path, calculate and determine the path space curvature index for each vulture individual. The path space curvature index is obtained by nonlinearly combining the cumulative value of the spatial gravitational gradient with the path length. The path length is defined as the number of data points on the spatiotemporal path.

[0073] S553. Based on the spatial curvature index of the path from each vulture individual to each gravity center node, calculate the topological coupling strength between the paths one by one. The topological coupling strength is defined as a nonlinear function of the difference in spatial curvature index between each path, and generate a path topological coupling matrix centered on the vulture individual.

[0074] S554. Based on the path topology coupling matrix, determine the global gravitational stability index of each path for each vulture individual. The global gravitational stability index is defined as the nonlinear weighted combination of the path space curvature index and the topology coupling strength.

[0075] S555. Based on the global gravitational stability index of the path, sort the paths from each vulture individual to all gravitational center nodes, and determine in real time the current target gravitational node of the vulture individual as the gravitational center node connected by the path with the largest global gravitational stability index.

[0076] S556. Monitor and record the changes in the current position coordinates and gravitational intensity of the characteristic space gravitational field of individual vultures in real time. When the position coordinates of individual vultures or the gravitational intensity in the characteristic space gravitational field change, recalculate and dynamically adjust the target gravitational nodes in real time to ensure that each individual vulture is always associated with the gravitational center node with the largest global gravitational stability index.

[0077] Optionally, S6 specifically includes:

[0078] S61. Based on the dynamic optimization sequence of model parameters, determine the range of cross-scale data overlap area corresponding to each optimization time in the sequence. The range of cross-scale data overlap area is determined based on the spatial location distribution trajectory sequence in the dynamic optimization sequence.

[0079] S62. Based on the position coordinates of each data point in the cross-scale data overlap area, calculate the back-diffusion influence factor of each data point in the back-diffusion reconstruction process. The back-diffusion influence factor is defined as a nonlinear combination of the position distribution probability of the data point in the optimization sequence and the rate of change of the numerical value in the cross-scale fused data matrix.

[0080] S63. Using the calculated inverse diffusion influence factor, dynamically adjust the data fusion weights during cross-scale data overlap and fusion point by point.

[0081] S64. Calculate the dynamic decrease rate of random noise weights for each data point in the cross-scale fused data matrix based on the dynamic optimization sequence of the model parameters.

[0082] S65. Based on the dynamically adjusted data fusion weight and the dynamic reduction rate of random noise weight, the cross-scale data overlap fusion and reverse diffusion process is executed point by point. The reverse diffusion process reconstructs the data scale by scale by gradually reducing random noise and tracing back the forward diffusion state.

[0083] S66. Record the data reconstruction results in the backdiffusion process in real time at each scale, and merge the recorded results at each scale to obtain the dynamic optimization spatiotemporal energy consumption behavior characteristic sequence after multi-scale backdiffusion reconstruction, which is updated in real time with the dynamic optimization sequence of model parameters.

[0084] The beneficial effects of this invention are:

[0085] (1) This invention combines a spatiotemporal diffusion probability generation model with a quantum chaotic African vulture optimization algorithm, utilizing the dynamic optimization sequence of the spatiotemporal data matrix and the inverse diffusion reconstruction process to overcome the limitations of existing energy consumption behavior simulation methods in handling dynamic and complex environments. This method effectively improves the accuracy of energy consumption prediction and enhances the model's adaptability to variable environments, especially exhibiting higher stability and prediction accuracy when facing complex, nonlinear, and dynamically changing energy consumption behaviors.

[0086] (2) This invention significantly improves the multi-scale fusion and optimization process of spatiotemporal data by introducing a dynamic noise weight reduction mechanism and a quantum chaotic African vulture optimization algorithm in the spatiotemporal diffusion probability generation model, thereby enhancing the adaptive capability in the data processing process. This technical measure enables the model to more accurately simulate and predict energy consumption behavior at different spatiotemporal scales, improves the model's generalization ability, and has better performance, especially in complex and dynamically changing energy consumption systems.

[0087] (3) In the field of energy consumption behavior simulation, this invention effectively solves the shortcomings of existing technologies in processing multi-dimensional features and their interactive effects of spatiotemporal data by introducing cross-scale data overlap fusion and dynamic optimization sequences. Through multi-scale optimization methods, this invention breaks through the bottlenecks of existing technologies in global and local optimization, especially in real-time dynamic data processing and multi-dimensional feature modeling, achieving significant technological progress and improving the reliability and accuracy of energy consumption prediction. Attached Figure Description

[0088] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0089] Figure 1 This is a flowchart illustrating the spatiotemporal data matrix construction process for an energy consumption behavior simulation method based on a generative model proposed in this invention.

[0090] Figure 2 This is a schematic diagram illustrating the optimization process of the quantum chaotic African vulture optimization algorithm for an energy consumption behavior simulation method based on a generative model proposed in this invention.

[0091] Figure 3 This is a flowchart of the reverse diffusion reconstruction and multi-scale data optimization process for an energy consumption behavior simulation method based on a generative model proposed in this invention. Detailed Implementation

[0092] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0093] refer to Figures 1-3 A method for simulating energy consumption behavior based on a generative model includes the following steps:

[0094] S1. Based on the overall continuity characteristics of the spatiotemporal energy consumption data of the collected equipment, the locations of data anomalies and missing data are dynamically determined point by point, and bidirectional interpolation reconstruction is performed. A high-dimensional continuous spatiotemporal data matrix is ​​generated through dynamic data scale transformation.

[0095] S2. Based on a high-dimensional continuous spatiotemporal data matrix, the data matrix is ​​alternately decomposed into multi-scale sub-matrices along the spatiotemporal axis. The forward diffusion process of each scale sub-matrix is ​​driven by random dynamic noise to obtain spatiotemporal scale multi-dimensional encoded data.

[0096] S3. Based on multi-dimensional encoded data, the reverse diffusion process is performed by using cross-scale data overlap fusion and dynamic noise weight reduction mechanism to realize the layer-by-layer restoration of encoded data at each scale under multi-scale fusion constraints, and obtain the reconstructed spatiotemporal energy consumption initial data.

[0097] S4. Based on the reconstructed initial spatiotemporal energy consumption data, construct the characteristic spatial gravitational field of the spatiotemporal energy consumption data, and determine the spatial location distribution of the initial quantum chaotic African vulture group.

[0098] S5. Within the gravitational field of the characteristic space, the positions of individual members of the quantum African vulture group are updated in a dynamic chaotic probability manner, so that the group positions are dynamically adjusted to adapt to the gravitational gradient of the spatiotemporal data, thereby obtaining a dynamic optimization sequence of model parameters.

[0099] S6. Utilizing the dynamic optimization sequence, the cross-scale data overlap fusion and dynamic noise weight reduction mechanism are adjusted in reverse to perform closed-loop adaptive optimization of the multi-scale reverse diffusion reconstruction process, thereby obtaining a dynamically optimized spatiotemporal energy consumption behavior characteristic sequence.

[0100] By combining a spatiotemporal diffusion probability generation model with a quantum chaotic African vulture optimization algorithm, this invention achieves accurate reconstruction and adaptive optimization of device energy consumption data. Specifically, this invention constructs a high-dimensional continuous spatiotemporal data matrix through dynamic data interpolation and scale transformation, and reconstructs the data using a random dynamic noise forward diffusion and cross-scale backward diffusion mechanism. Simultaneously, by utilizing the characteristic space gravitational field construction and a quantum chaotic optimization algorithm, the model can effectively handle multi-scale fusion and parameter optimization of complex dynamic data. This method exhibits high accuracy and generalization ability when dealing with dynamically changing spatiotemporal energy consumption behavior, overcoming the shortcomings of existing technologies in predicting accuracy and adaptability in complex environments.

[0101] In this embodiment, S1 specifically includes:

[0102] S11. Based on the raw spatiotemporal energy consumption data collected by the equipment, arrange the data in the continuous order of the data collection time to form an initial data sequence, calculate the local density value between each data point and the data points at multiple times in its neighborhood, and dynamically determine the location of data anomalies in the initial data sequence based on the gradient change of the local density value over time.

[0103] S12. Based on the location of outliers, the initial data sequence is divided into several continuous data segments with adjacent outlier locations as the dividing boundaries. The integrity of each data segment is evaluated separately. Based on the sampling time interval of the data points in the time dimension, the specific location of the missing data points within the segment is dynamically determined.

[0104] S13. Using the missing data locations as interpolation targets, establish a spatiotemporal neighborhood window centered on the missing data points, and calculate the information entropy values ​​of adjacent data points within each neighborhood window. By gradually minimizing the information entropy differences among data points within the neighborhood, determine the interpolation weights of the interpolation target points in both spatiotemporal dimensions, perform bidirectional adaptive interpolation reconstruction, and generate reconstructed data values ​​for each missing location.

[0105] S14. After repairing the missing positions of each data sub-segment with reconstructed data values, a temporally continuous sub-segment overlap area is formed by overlapping sub-segment boundaries. Data smoothing weights are established through the numerical differences between data points in the boundary overlap area. Adaptive numerical smoothing fusion of the boundary area is performed on each sub-segment to keep the data values ​​between adjacent sub-segments continuous and form a continuous spatiotemporal data sequence.

[0106] S15. Using continuous spatiotemporal data sequences, a two-dimensional spatiotemporal data matrix is ​​constructed by mapping the data points in the data sequence to spatiotemporal two-dimensional coordinates. The numerical distribution imbalance of the data points in the data matrix is ​​evaluated in a local region. The dynamic conversion weight coefficient of the data scale is determined based on the ratio of the numerical variance in the local region to the mean of the neighborhood. The numerical scale of the data points in the data matrix is ​​dynamically adjusted region by region.

[0107] S16. Based on the scaled data matrix, perform high-dimensional nonlinear spatial expansion mapping of the data row by row and column by column. Using the row and column positions of each data point in the original matrix as the mapping basis, generate multiple high-dimensional subspace mapping data points through nonlinear mapping functions, and recombine the mapping data points to form a high-dimensional continuous spatiotemporal data matrix.

[0108] By meticulously identifying and adaptively interpolating outliers and missing data in spatiotemporal energy consumption data, and further utilizing adaptive numerical smoothing fusion and dynamic scaling transformation of overlapping sub-segments, this invention achieves accurate construction of a high-dimensional continuous spatiotemporal data matrix. Compared to existing single interpolation and simple data fusion methods, this invention can more accurately restore the data continuity in missing and outlier regions, effectively improving the completeness and accuracy of data reconstruction.

[0109] In this embodiment, S2 specifically includes:

[0110] S21. Based on the spatiotemporal coordinates of each data point in the high-dimensional continuous spatiotemporal data matrix, a progressive non-uniform density threshold segmentation strategy is adopted along the time axis and the spatial axis to calculate the spatiotemporal data density difference in the local region of the matrix. Based on the gradient change of density difference, the matrix is ​​dynamically decomposed into multiple data sub-matrices with progressively changing scale and irregular boundaries.

[0111] S22. Based on each data submatrix, calculate the spatial correlation coefficient of each data point within the submatrix relative to its local spatiotemporal neighbor data points, and determine the spatial clustering center position of the data within the submatrix based on the spatial correlation coefficient, and generate submatrix spatial clustering maps that reflect the spatial distribution structure of the data one by one.

[0112] S23. Based on the cluster center positions of each submatrix spatial clustering mapping and the spatial correlation coefficients of each data point, the normalization scaling factor is determined point by point. The normalization scaling factor is determined by the nonlinear function of the spatial correlation coefficients between the data points and the cluster center positions. The submatrix is ​​spatially adaptively scaled and normalized in a point-by-point adjustment manner to obtain the scaled normalized data submatrix.

[0113] S24. Utilize the data density distribution within the data submatrix after scale normalization, dynamically determine the spatial distribution of random noise variance by the change in the spatial density gradient of the local neighborhood where each data point is located, and generate a dynamic random noise data matrix with adaptive variance based on the local data density gradient, so that the dynamic random noise within each submatrix exhibits a non-uniform distribution in the spatial dimension.

[0114] S25. The dynamic random noise data matrix is ​​superimposed one by one on the corresponding scale-normalized data submatrix. Based on the superimposed data submatrix, the local spatial density difference of the data points is calculated point by point. The diffusion intensity factor in the forward diffusion process is determined according to the spatial density difference. The diffusion degree of each data point is gradually adjusted, and the stepwise forward diffusion process between data points is executed.

[0115] S26. During the forward diffusion process, based on the diffusion state change process of data points in the data submatrix at each scale, the data density change trajectory of each data point in the local spatiotemporal neighborhood during the diffusion process is sampled and recorded in real time. Based on the data point density change trajectory, a multidimensional data encoding sequence containing spatiotemporal scale and diffusion state is generated to obtain spatiotemporal scale multidimensional encoded data.

[0116] By performing non-uniform density thresholding, spatial correlation clustering mapping, and adaptive scale normalization on a high-dimensional continuous spatiotemporal data matrix, and combining this with a dynamic random noise data matrix and a multi-scale forward diffusion process driven by spatial density differences, spatiotemporal scale multidimensional encoded data is obtained. Compared with existing methods of simple or static scale segmentation and uniform noise superposition, this invention effectively improves the adaptability and scale fusion accuracy in spatiotemporal data processing, enhances the model's ability to capture complex spatiotemporal data features, and provides a more stable data encoding foundation for accurately simulating complex energy consumption behavior.

[0117] In this embodiment, S3 specifically includes:

[0118] S31. Based on the spatiotemporal scale multidimensional encoded data matrix, calculate the spatial position offset of all data points in each scale encoded data matrix relative to the global data matrix, and determine the scale correlation strength of each data point based on the spatial position offset.

[0119] S32. Determine the overlapping area of ​​cross-scale data based on the scale correlation strength, and calculate the numerical coherence degree of corresponding data points between different scale coded data matrices in the overlapping area one by one. The numerical coherence degree is a nonlinear coherence function of the numerical difference of data points.

[0120] S33. Construct a data fusion path point by point based on the numerical synergy degree, taking the data point with the highest numerical synergy degree as the starting point of the fusion path, and gradually extending to the data points with lower numerical synergy degree, dynamically constructing a data fusion tree topology structure.

[0121] S34. Utilize the data fusion tree topology structure to perform cross-scale numerical fusion of data nodes in the tree topology structure one by one, so as to realize the step-by-step data fusion of data matrices of different scales under the fusion topology structure and form a cross-scale fusion data matrix.

[0122] S35. Based on the cross-scale fusion data matrix, calculate the local stability index of each data point and its spatial neighborhood, and dynamically generate a random noise weight matrix with non-uniform spatial distribution based on the gradient of the local stability index.

[0123] S36. The random noise weight matrix is ​​superimposed point by point onto the cross-scale fused data matrix. Based on the local stability index gradient of each data point in the superimposed matrix, the adaptive deceleration rate of the back diffusion intensity is determined point by point to dynamically adjust the gradual decrease process of the random noise weight.

[0124] S37. Based on the gradual decrease of random noise weight, reverse diffusion is performed point by point. The data fusion process is traced back layer by layer based on the fused data topology path. The initial values ​​before data diffusion are restored point by point, and finally the reconstructed spatiotemporal energy consumption initial data is obtained.

[0125] By constructing a tree-like data fusion topology, accurate fusion of data at different spatiotemporal scales is achieved. Furthermore, by dynamically determining noise weights and adaptive deceleration rates using a local stability index, the accuracy and stability of the inverse diffusion reconstruction process are effectively improved. Compared to simple linear or static fusion methods in existing technologies, the data fusion path and tree-like topology employed in this invention can more accurately capture subtle differences and correlations between data. The non-uniform spatial distribution of dynamic noise weights effectively reduces errors in the inverse diffusion process, providing a solid data foundation for high-precision prediction of complex energy consumption behavior.

[0126] In this embodiment, S33 specifically includes:

[0127] S331. For each data point in the cross-scale data overlap area, take the data point as the reference point and calculate the spatial topological distance between the reference point and all other data points in the area. The spatial topological distance is defined as a nonlinear weighted combination function of the Euclidean distance of spatiotemporal coordinates and the difference in numerical coherence.

[0128] S332. Using the calculated spatial topological distance, determine the fusion topological centrality index of each data point in the cross-scale data overlap region one by one. The fusion topological centrality index is obtained by non-linear weighted summation of the reciprocal of the spatial topological distance between the data point and all other data points in the region.

[0129] S333. Based on the fusion topology centrality index, perform global sorting and dynamically determine the data point with the highest fusion topology centrality index as the starting node of the fusion topology structure.

[0130] S334. Taking the starting node of the fused topology as the origin, the topological diffusion potential energy from the starting node to its spatial neighborhood data points is calculated step by step. The topological diffusion potential energy is obtained by a nonlinear combination function of the fused topological centrality index and the spatial topological distance, and the data point with the largest topological diffusion potential energy is taken as the next extension node of the fused path.

[0131] S335. Repeat the calculation of topological diffusion potential energy for each determined extension node, and gradually determine the next extension node based on the real-time change of topological diffusion potential energy. Dynamically adjust the topological extension direction of the fusion path and gradually generate multiple branch structure data fusion paths in a nonlinear manner.

[0132] S336. The fusion order of each node in the real-time computing data fusion path structure is a nonlinear sorting function that combines the node's topological diffusion potential energy and topological centrality index. The hierarchical position of each node on the data fusion path is determined based on the fusion order.

[0133] S337. Based on the determined data fusion path and fusion order, connect the topological relationships between data nodes level by level, establish a data fusion tree network with asymmetric topological branch structure point by point, and obtain a refined, nonlinear and multi-branch structure of cross-scale data fusion tree topology.

[0134] By introducing a nonlinear combination function of spatial topological distance and numerical synergy, a data fusion path construction method with fusion topological centrality as the core indicator is constructed, enabling global scheduling and multi-level path expansion of information within overlapping data regions across scales. Simultaneously, topological dissipation potential energy is used to dynamically adjust the extension direction of data nodes, combined with path hierarchy fusion depth sorting, to construct a data fusion tree network with a refined topological organization structure. Compared with existing technologies, this invention can more fully express the topological hierarchy and synergy characteristics between cross-scale data while maintaining structural coherence, thus improving the spatial representation capability of multi-scale information fusion in energy consumption behavior simulation.

[0135] In this embodiment, S35 specifically includes:

[0136] S351. Taking each data point in the cross-scale fused data matrix as the center, determine the spatial neighborhood window centered on the data point. The size of the neighborhood window is a fixed spatial radius range centered on the data point.

[0137] S352. Calculate the difference between the data point values ​​in the spatial neighborhood window and the center data point value one by one, and determine the degree of numerical deviation between the center data point and the spatial neighborhood data points based on the difference.

[0138] S353. Based on the degree of numerical deviation of all data points within the neighborhood window, determine the local stability index L of the central data point point by point. SI :

[0139]

[0140] Among them, L SI (i,j) is the local stability index of the central data point at position (i,j) within the fused data matrix, Ω ij Let N be the set of data points within the spatial neighborhood window of the center data point at position (i,j). ij V(i,j) represents the total number of data points within the spatial neighborhood window, and V(m,n) represents the values ​​of the center data point at position (i,j) and the neighborhood data point at position (m,n), respectively.

[0141] S354. Based on the local stability index of all data points in the cross-scale fused data matrix, calculate the stability index gradient of each data point in the spatial dimension. The stability index gradient is defined as the spatial weighted average of the difference between the local stability index of each data point and the local stability index of its neighboring data points.

[0142] S355. Based on the stability index gradient of the data points, determine the spatial distribution intensity of the random noise weights point by point:

[0143]

[0144] Among them, W ij Let (i,j) be the random noise weight for the data point at position (i,j). Let be the absolute value of the stability exponent gradient of the data point at position (i,j). To maximize the absolute value of the stability exponent gradient of all data points within the data matrix, W max This is the preset maximum intensity value for the random noise weights;

[0145] S356. Based on the spatial distribution intensity of the calculated random noise weight, generate random noise values ​​point by point. The random noise values ​​follow a normal distribution with the random noise weight as the standard deviation and the mean as zero.

[0146] S357. Based on the cross-scale fusion data matrix, the generated random noise values ​​are superimposed point by point onto the corresponding data points to finally generate a random noise weight matrix with a non-uniform spatial distribution.

[0147] By introducing a dynamic collaborative calculation mechanism of local stability index and random noise weights, spatial adaptive control of noise perturbation intensity is achieved during multi-scale data fusion. Specifically, this invention utilizes the numerical deviation of data points within their local neighborhoods to construct a refined stability evaluation index, and generates non-uniformly distributed random noise weights based on the gradient of this index, thereby improving the targeting and resolution of noise perturbations at each data point during reconstruction. Compared with the uniform noise perturbation processing method in existing technologies, this invention significantly improves the spatial smoothness and local structure preservation ability of reconstructed data, enhancing the stability and accuracy of the model when dealing with dynamic energy consumption behavior.

[0148] In this embodiment, S4 specifically includes:

[0149] S41. Based on the reconstructed initial spatiotemporal energy consumption data matrix, calculate the nonlinear combination of spatiotemporal distance and numerical difference between each data point and other data points in the matrix point by point, so as to determine the characteristic spatial potential energy value of each data point.

[0150] S42. Based on the characteristic spatial potential energy value of each data point, determine the potential energy spatial gradient vector of each position in the data matrix one by one through the spatial position coordinates. The magnitude and direction of the spatial gradient vector are clearly determined by the spatial partial derivative of the characteristic spatial potential energy value.

[0151] S43. Based on the potential energy spatial gradient vector at each position in the data matrix, calculate and determine the spatial derivative of the potential energy gradient vector point by point to generate the spatial rate of change matrix of the potential energy gradient vector.

[0152] S44. Based on the potential energy gradient vector space rate of change matrix, calculate the characteristic space gravitational intensity at each position in the data matrix space. The gravitational intensity value is determined by the absolute value of the value at each position in the potential energy gradient vector space rate of change matrix.

[0153] S45. Based on all the gravitational intensities of the characteristic space determined in the data matrix, construct a clear characteristic space gravitational field point by point. The field strength at each position in the characteristic space gravitational field is directly represented by the numerical value of the characteristic space gravitational intensity at the corresponding position.

[0154] S46. Based on the gravitational intensity distribution characteristics within the characteristic space gravitational field, and taking the relative magnitude of the local gravitational intensity as the basis, determine several spatial gravitational center nodes in the gravitational field. The gravitational center node is defined as the location point where the local gravitational intensity is extremely large in the characteristic space gravitational field.

[0155] S47. Based on the determined position of the spatial gravitational center node and the gravitational field strength, dynamically determine the spatial position distribution of the initial quantum chaotic African vulture group, and determine the specific spatial position coordinates of the initial quantum chaotic African vulture group by non-uniform random sampling according to the position distribution.

[0156] By constructing a gravitational field in the characteristic space and calculating the potential energy and gradient vector of the characteristic space based on the nonlinear spatiotemporal distance combination function between data points, a refined modeling of the gravitational relationship among data points in the initial energy consumption data was achieved. Based on this, a gravitational intensity distribution matrix was further generated, and the gravitational center was precisely defined according to the gravitational gradient direction and the location of local maxima, thus constructing a gravitational field model with spatially driven characteristics. Finally, the initial position layout of the initial quantum chaotic African vulture colony in the characteristic space was completed through a non-uniform probability sampling strategy. Compared with the random or equally spaced initialization methods used in existing technologies, this scheme can more effectively guide the optimization individuals to focus on key regions, improving the efficiency and spatial convergence capability of the model in the global search phase.

[0157] In this embodiment, S5 specifically includes:

[0158] S51. Based on the spatial position coordinates of each individual in the initial quantum chaotic African vulture group in the characteristic space gravitational field, calculate the spatiotemporal gravitational gradient vector between each vulture individual and all gravitational center nodes in the gravitational field, and form the global gravitational perception vector of the vulture individual.

[0159] S52. Based on the global gravitational perception vector of an individual vulture, the local chaotic perturbation probability at the current position is calculated in real time for each individual vulture. The local chaotic perturbation probability is determined by the nonlinear weighted sum of the change rates of the gravitational gradient vector at each position in the gravitational field of the characteristic space.

[0160] S53. Based on the local chaotic perturbation probability, dynamically determine the perturbation threshold of the position update direction for each vulture individual, and use the perturbation threshold to adjust the position update step size of the vulture individual along the gravitational gradient vector direction in real time.

[0161] S54. Calculate the dynamic spatial topological distance between the current spatial position of each vulture individual and the positions of each gravitational center node in the characteristic spatial gravitational field. The dynamic spatial topological distance considers a combination of Euclidean distance of position coordinates and nonlinear function of gravitational intensity difference.

[0162] S55. Based on the calculated dynamic spatial topological distance, the target gravitational node of each vulture individual is dynamically determined. The target gravitational node is adjusted in real time and defined as the gravitational center node with the smallest calculated value of the combination of topological distance and gravitational intensity.

[0163] S56. Each vulture individual is updated with its spatial coordinates gradually through a perturbation threshold, based on a dynamically determined target gravity node. The update method is constrained in real time by the combined effect of the gravity gradient vector and the probability of local chaotic perturbation.

[0164] S57. Record the updated spatial coordinates of each vulture individual in real time, form a trajectory sequence for dynamic adjustment of the group position one by one, and finally obtain the vulture group position distribution trajectory sequence that dynamically changes with the characteristic spatial gravitational field, and determine the position distribution trajectory sequence as the dynamic optimization sequence of model parameters.

[0165] By constructing the relationship between the position and gravitational gradient of an individual member of a quantum chaotic African vulture colony in a characteristic spatial gravitational field, and combining this with a local chaotic perturbation probability mechanism, the update trajectory of each individual along the gravitational gradient direction is dynamically adjusted, achieving a unified modeling of the global and local search behavior of the optimized individuals. Furthermore, by introducing a combined calculation method of dynamic spatial topological distance and nonlinear gravitational intensity function, the accuracy and scalability of the direction selection for individual position updates are further improved. This mechanism can effectively adapt to the differences in gradient distribution of spatiotemporal energy consumption characteristic data in different regions, improving the convergence stability and search balance of the optimization algorithm in a dynamic environment, and providing a more structured solution space navigation strategy for subsequent dynamic optimization of model parameters.

[0166] In this embodiment, S55 specifically includes:

[0167] S551. For each vulture individual's current position, construct multiple spatiotemporal paths from the vulture individual to each gravitational center node in the characteristic spatial gravitational field, and calculate the cumulative value of the spatial gravitational gradient of all data points on each path. The cumulative value of the spatial gravitational gradient is the point-by-point summation of the absolute values ​​of the gravitational gradient at each data point on the path.

[0168] S552. Based on the cumulative value of the spatial gravitational gradient of each spatiotemporal path, calculate and determine the path space curvature index for each vulture individual. The path space curvature index is obtained by nonlinearly combining the cumulative value of the spatial gravitational gradient with the path length. The path length is defined as the number of data points on the spatiotemporal path.

[0169] S553. Based on the spatial curvature index of the path from each vulture individual to each gravity center node, calculate the topological coupling strength between the paths one by one. The topological coupling strength is defined as a nonlinear function of the difference in spatial curvature index between each path, and generate a path topological coupling matrix centered on the vulture individual.

[0170] S554. Based on the path topology coupling matrix, determine the global gravitational stability index of each path for each vulture individual. The global gravitational stability index is defined as the nonlinear weighted combination of the path space curvature index and the topology coupling strength.

[0171] S555. Based on the global gravitational stability index of the path, sort the paths from each vulture individual to all gravitational center nodes, and determine in real time the current target gravitational node of the vulture individual as the gravitational center node connected by the path with the largest global gravitational stability index.

[0172] S556. Monitor and record the changes in the current position coordinates and gravitational intensity of the characteristic space gravitational field of individual vultures in real time. When the position coordinates of individual vultures or the gravitational intensity in the characteristic space gravitational field change, recalculate and dynamically adjust the target gravitational nodes in real time to ensure that each individual vulture is always associated with the gravitational center node with the largest global gravitational stability index.

[0173] By introducing a spatial path curvature modeling mechanism based on gravitational gradient accumulation, and combining a nonlinear weighted fusion of global gravitational stability indices and path topological coupling, an adaptive gravitational path selection strategy for individual quantum chaotic African vultures is constructed. Compared to existing techniques that rely on static Euclidean distance or local shortest paths for path selection, this method can dynamically adjust the path and direction of the individual's migration to the gravitational center node, improving the global guidance capability and path stability during the search process. This allows for more effective guidance of individuals to quickly converge to the global target region in high-dimensional space, enhancing the stability and structural continuity of the optimized path.

[0174] In this embodiment, S6 specifically includes:

[0175] S61. Based on the dynamic optimization sequence of model parameters, determine the range of cross-scale data overlap area corresponding to each optimization time in the sequence. The range of cross-scale data overlap area is determined based on the spatial location distribution trajectory sequence in the dynamic optimization sequence.

[0176] S62. Based on the position coordinates of each data point in the cross-scale data overlap area, calculate the back-diffusion influence factor of each data point in the back-diffusion reconstruction process. The back-diffusion influence factor is defined as a nonlinear combination of the position distribution probability of the data point in the optimization sequence and the rate of change of the numerical value in the cross-scale fused data matrix.

[0177] S63. Using the calculated inverse diffusion influence factor, dynamically adjust the data fusion weights point-by-point during cross-scale data overlap fusion:

[0178]

[0179] Where, β ij (t) represents the data fusion weight at data point position (i,j) at optimization time t, β0 is the initial data fusion weight constant, and P ij (t) represents the dynamic distribution probability of the data point position (i,j) in the optimized sequence, R ij(t) represents the rate of change of the data value at position (i,j). This represents the maximum value of the product of the dynamic distribution probability and the rate of change of the value among all data points;

[0180] S64. Calculate the dynamic decrease rate of random noise weights for each data point in the cross-scale fused data matrix based on the dynamic optimization sequence of the model parameters:

[0181]

[0182] Where, γ ij (t) represents the dynamic decrease rate of the random noise weight at the data point position (i,j) at optimization time t, γ max L is the maximum deceleration rate constant for random noise weights, k is the deceleration rate adjustment constant, and L is the maximum deceleration rate constant for random noise weights. SI (i,j,t) represents the local stability index at position (i,j) at time t during optimization. To optimize the maximum local stability exponent among all data points at time t;

[0183] S65. Based on the dynamically adjusted data fusion weight and the dynamic reduction rate of random noise weight, the cross-scale data overlap fusion and reverse diffusion process is executed point by point. The reverse diffusion process reconstructs the data scale by scale by gradually reducing random noise and tracing back the forward diffusion state.

[0184] S66. Record the data reconstruction results in the backdiffusion process in real time at each scale, and merge the recorded results at each scale to obtain the dynamic optimization spatiotemporal energy consumption behavior characteristic sequence after multi-scale backdiffusion reconstruction, which is updated in real time with the dynamic optimization sequence of model parameters.

[0185] By constructing a joint regulation mechanism based on a dynamically optimized sequence-driven inverse diffusion data fusion weighting factor and a random noise reduction factor, dynamic adaptive adjustment of the fusion accuracy and perturbation intensity of each data point during multi-scale inverse diffusion reconstruction is achieved. Specifically, the fusion weighting factor is generated based on the nonlinear correlation between the dynamic distribution probability and the rate of change of the data points, while the noise reduction factor controls the attenuation of perturbation intensity by combining the changing trend of the local stability index. Compared with the fixed or uniform processing of weights and perturbation mechanisms in existing technologies, this scheme can respond more sensitively to changes in data characteristics in dynamic scenarios, improve the accuracy and stability of multi-scale data inverse reconstruction, and provide a continuous and consistent optimized data foundation for the characteristic representation of spatiotemporal energy consumption behavior.

[0186] Example 1:

[0187] To verify the feasibility of this invention in practice, it was applied to the industrial plant energy consumption simulation platform, which is a key component of the energy management system in a certain region. Spatiotemporal dynamic modeling and energy consumption behavior prediction were performed on the real-time power consumption data of different industrial equipment in the region collected at high frequency. The goal is to simulate and evaluate the electricity consumption trend for the next 12 hours, so as to optimize energy allocation strategies and improve the accuracy of load forecasting.

[0188] In this application scenario, traditional energy consumption simulation methods mainly rely on statistical regression, static clustering, or machine learning modeling to predict trends by fitting models to historical data. These methods generally suffer from static model parameters and a lack of adaptive optimization capabilities. Furthermore, they struggle to depict the complex spatiotemporal nonlinear interactions between devices caused by factors such as spatial layout and changes in operating load, resulting in poor model generalization ability. When faced with sudden load fluctuations, adjustments to energy-saving strategies, or coordinated scheduling across different regions, traditional methods exhibit high prediction errors and slow response times, making it difficult to meet the processing needs of large-scale, multi-source, heterogeneous energy consumption data.

[0189] In practical application, this invention first collects three days of spatiotemporal distributed energy consumption data from 12 types of industrial equipment within a factory area, with a sampling frequency of once every 5 minutes, resulting in a total data volume exceeding 100,000 records. The system constructs a continuous spatiotemporal data matrix from the raw data, utilizes a dynamic noise-driven mechanism and scale normalization strategy to handle missing data and outliers, and constructs a high-dimensional continuous spatiotemporal energy consumption data structure through bidirectional interpolation and local smoothing fusion. Subsequently, a spatiotemporal diffusion probability generation model is used to encode the data at multiple scales, and an inverse diffusion mechanism with dynamically decreasing weights is combined to reconstruct initial energy consumption behavior features with high feature retention.

[0190] After feature extraction, the system calculates gravitational potential energy and gradient based on the feature space gravitational field construction mechanism, generates a non-uniform gravitational field, and dynamically samples the positions of individual members in the initial quantum chaotic African vulture colony. During the optimization phase, the vulture individuals, driven by local chaotic perturbation probabilities, adaptively iterate along the gravitational gradient direction, gradually converging to the high-density region of the feature energy consumption trajectory. In the model optimization process, the dynamic position adjustment process is dynamically bound to the inverse diffusion parameters and fusion path, achieving closed-loop adaptive iteration.

[0191] In the performance evaluation, we selected energy consumption data from 52 typical devices on the platform and compared the method of this invention with traditional LSTM (Long Short-Term Memory) and XGBoost (Gradient Boosting Tree) methods. The evaluation metrics included mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). The test results are shown in the table below.

[0192] Table 1. Performance Comparison of Multi-Model in Different Energy Prediction Tasks

[0193]

[0194] As can be seen from Table 1, among all the test devices, the model constructed by the present invention outperforms the comparative methods in the three core indicators of MAE, RMSE, and MAPE. Taking device D-033 as an example, the MAE of its prediction using the model of the present invention is 0.597, which is significantly lower than 1.010 of LSTM and 1.122 of XGBoost. At the same time, its MAPE is controlled at 2.63%, reducing the error range by nearly 40% compared with the traditional model, and significantly improving the prediction accuracy.

[0195] Further analysis reveals that the model of the present invention still maintains stable prediction performance when dealing with data mutation points, periodic variations, and non-stationary operating loads. In a sudden load adjustment, the system automatically adjusts the inverse diffusion path parameters based on the dynamic optimization trajectory, and the error recovery rate reaches 67.2%, verifying that the method has strong robustness and adaptability. In addition, the system can continuously update the fusion path structure during the deployment process and fine-tune and optimize according to the latest feedback samples, realizing the online linkage evolution of the model structure and parameters.

[0196] In summary, the energy consumption behavior simulation method proposed by the present invention has significant advantages compared with traditional methods in terms of spatio-temporal structure modeling, dynamic optimization, and prediction accuracy control. It is especially suitable for application environments with complex industrial equipment operation states and剧烈的能耗动态变化剧烈的, providing high-performance and scalable technical support for industrial-level energy consumption prediction systems.

[0197] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A method for simulating energy consumption behavior based on a generative model, characterized in that, Includes the following steps: S1. Based on the overall continuity characteristics of the spatiotemporal energy consumption data of the collected equipment, the locations of data anomalies and missing data are dynamically determined point by point, and bidirectional interpolation reconstruction is performed. A high-dimensional continuous spatiotemporal data matrix is ​​generated through dynamic data scale transformation. S2. Based on a high-dimensional continuous spatiotemporal data matrix, the data matrix is ​​alternately decomposed into multi-scale sub-matrices along the spatiotemporal axis. The forward diffusion process of each scale sub-matrix is ​​driven by random dynamic noise to obtain spatiotemporal scale multidimensional encoded data. S3. Based on multi-dimensional encoded data, the reverse diffusion process is performed by using cross-scale data overlap fusion and dynamic noise weight reduction mechanism to realize the layer-by-layer restoration of encoded data at each scale under multi-scale fusion constraints, and obtain the reconstructed spatiotemporal energy consumption initial data. S4. Based on the reconstructed initial spatiotemporal energy consumption data, construct the characteristic spatial gravitational field of the spatiotemporal energy consumption data, and determine the spatial location distribution of the initial quantum chaotic African vulture group. S5. Within the gravitational field of the characteristic space, the positions of individual members of the quantum African vulture group are updated in a dynamic chaotic probability manner, so that the group positions are dynamically adjusted to adapt to the gravitational gradient of the spatiotemporal data, thereby obtaining a dynamic optimization sequence of model parameters. S6. Utilizing the dynamic optimization sequence, the cross-scale data overlap fusion and dynamic noise weight reduction mechanism are adjusted in reverse to perform closed-loop adaptive optimization of the multi-scale reverse diffusion reconstruction process, thereby obtaining a dynamically optimized spatiotemporal energy consumption behavior characteristic sequence.

2. The energy consumption behavior simulation method based on a generative model according to claim 1, characterized in that, S1 specifically includes: S11. Based on the raw spatiotemporal energy consumption data collected by the equipment, arrange the data in the continuous order of the data collection time to form an initial data sequence, calculate the local density value between each data point and the data points at multiple times in its neighborhood, and dynamically determine the location of data anomalies in the initial data sequence based on the gradient change of the local density value over time. S12. Based on the location of outliers, the initial data sequence is divided into several continuous data segments with adjacent outlier locations as the dividing boundaries. The integrity of each data segment is evaluated separately. Based on the sampling time interval of the data points in the time dimension, the specific location of the missing data points within the segment is dynamically determined. S13. Using the missing data locations as interpolation targets, establish a spatiotemporal neighborhood window centered on the missing data points, and calculate the information entropy values ​​of adjacent data points within each neighborhood window. By gradually minimizing the information entropy differences among data points within the neighborhood, determine the interpolation weights of the interpolation target points in both spatiotemporal dimensions, perform bidirectional adaptive interpolation reconstruction, and generate reconstructed data values ​​for each missing location. S14. After repairing the missing positions of each data sub-segment with reconstructed data values, a temporally continuous sub-segment overlap area is formed by overlapping sub-segment boundaries. Data smoothing weights are established through the numerical differences between data points in the boundary overlap area. Adaptive numerical smoothing fusion of the boundary area is performed on each sub-segment to keep the data values ​​between adjacent sub-segments continuous and form a continuous spatiotemporal data sequence. S15. Using continuous spatiotemporal data sequences, a two-dimensional spatiotemporal data matrix is ​​constructed by mapping the data points in the data sequence to spatiotemporal two-dimensional coordinates. The numerical distribution imbalance of the data points in the data matrix is ​​evaluated in a local region. The dynamic conversion weight coefficient of the data scale is determined based on the ratio of the numerical variance in the local region to the mean of the neighborhood. The numerical scale of the data points in the data matrix is ​​dynamically adjusted region by region. S16. Based on the scaled data matrix, perform high-dimensional nonlinear spatial expansion mapping of the data row by row and column by column. Using the row and column positions of each data point in the original matrix as the mapping basis, generate multiple high-dimensional subspace mapping data points through nonlinear mapping functions, and recombine the mapping data points to form a high-dimensional continuous spatiotemporal data matrix.

3. The energy consumption behavior simulation method based on a generative model according to claim 1, characterized in that, S2 specifically includes: S21. Based on the spatiotemporal coordinates of each data point in the high-dimensional continuous spatiotemporal data matrix, a progressive non-uniform density threshold segmentation strategy is adopted along the time axis and the spatial axis to calculate the spatiotemporal data density difference in the local region of the matrix. Based on the gradient change of density difference, the matrix is ​​dynamically decomposed into multiple data sub-matrices with progressively changing scale and irregular boundaries. S22. Based on each data submatrix, calculate the spatial correlation coefficient of each data point within the submatrix relative to its local spatiotemporal neighbor data points, and determine the spatial clustering center position of the data within the submatrix based on the spatial correlation coefficient, and generate submatrix spatial clustering maps that reflect the spatial distribution structure of the data one by one. S23. Based on the cluster center positions of each submatrix spatial clustering mapping and the spatial correlation coefficients of each data point, the normalization scaling factor is determined point by point. The normalization scaling factor is determined by the nonlinear function of the spatial correlation coefficients between the data points and the cluster center positions. The submatrix is ​​spatially adaptively scaled and normalized in a point-by-point adjustment manner to obtain the scaled normalized data submatrix. S24. Utilize the data density distribution within the data submatrix after scale normalization, dynamically determine the spatial distribution of random noise variance by the change in the spatial density gradient of the local neighborhood where each data point is located, and generate a dynamic random noise data matrix with adaptive variance based on the local data density gradient, so that the dynamic random noise within each submatrix exhibits a non-uniform distribution in the spatial dimension. S25. The dynamic random noise data matrix is ​​superimposed one by one on the corresponding scale-normalized data submatrix. Based on the superimposed data submatrix, the local spatial density difference of the data points is calculated point by point. The diffusion intensity factor in the forward diffusion process is determined according to the spatial density difference. The diffusion degree of each data point is gradually adjusted, and the stepwise forward diffusion process between data points is executed. S26. During the forward diffusion process, based on the diffusion state change process of data points in the data submatrix at each scale, the data density change trajectory of each data point in the local spatiotemporal neighborhood during the diffusion process is sampled and recorded in real time. Based on the data point density change trajectory, a multidimensional data encoding sequence containing spatiotemporal scale and diffusion state is generated to obtain spatiotemporal scale multidimensional encoded data.

4. The energy consumption behavior simulation method based on a generative model according to claim 1, characterized in that, S3 specifically includes: S31. Based on the spatiotemporal scale multidimensional encoded data matrix, calculate the spatial position offset of all data points in each scale encoded data matrix relative to the global data matrix, and determine the scale correlation strength of each data point based on the spatial position offset. S32. Determine the overlapping area of ​​cross-scale data based on the scale correlation strength, and calculate the numerical coherence degree of corresponding data points between different scale coded data matrices in the overlapping area one by one. The numerical coherence degree is a nonlinear coherence function of the numerical difference of data points. S33. Construct a data fusion path point by point based on the numerical synergy degree, taking the data point with the highest numerical synergy degree as the starting point of the fusion path, and gradually extending to the data points with lower numerical synergy degree, dynamically constructing a data fusion tree topology structure. S34. Utilize the data fusion tree topology structure to perform cross-scale numerical fusion of data nodes in the tree topology structure one by one, so as to realize the step-by-step data fusion of data matrices of different scales under the fusion topology structure and form a cross-scale fusion data matrix. S35. Based on the cross-scale fusion data matrix, calculate the local stability index of each data point and its spatial neighborhood, and dynamically generate a random noise weight matrix with non-uniform spatial distribution based on the gradient of the local stability index. S36. The random noise weight matrix is ​​superimposed point by point onto the cross-scale fused data matrix. Based on the local stability index gradient of each data point in the superimposed matrix, the adaptive deceleration rate of the back diffusion intensity is determined point by point to dynamically adjust the gradual decrease process of the random noise weight. S37. Based on the gradual decrease of random noise weight, reverse diffusion is performed point by point. The data fusion process is traced back layer by layer based on the fused data topology path. The initial values ​​before data diffusion are restored point by point, and finally the reconstructed spatiotemporal energy consumption initial data is obtained.

5. The energy consumption behavior simulation method based on a generative model according to claim 4, characterized in that, S33 specifically includes: S331. For each data point in the cross-scale data overlap area, take the data point as the reference point and calculate the spatial topological distance between the reference point and all other data points in the area. The spatial topological distance is defined as a nonlinear weighted combination function of the Euclidean distance of spatiotemporal coordinates and the difference in numerical coherence. S332. Using the calculated spatial topological distance, determine the fusion topological centrality index of each data point in the cross-scale data overlap region one by one. The fusion topological centrality index is obtained by non-linear weighted summation of the reciprocal of the spatial topological distance between the data point and all other data points in the region. S333. Based on the fusion topology centrality index, perform global sorting and dynamically determine the data point with the highest fusion topology centrality index as the starting node of the fusion topology structure. S334. Taking the starting node of the fused topology as the origin, the topological diffusion potential energy from the starting node to its spatial neighborhood data points is calculated step by step. The topological diffusion potential energy is obtained by a nonlinear combination function of the fused topological centrality index and the spatial topological distance, and the data point with the largest topological diffusion potential energy is taken as the next extension node of the fused path. S335. Repeat the calculation of topological diffusion potential energy for each determined extension node, and gradually determine the next extension node based on the real-time change of topological diffusion potential energy. Dynamically adjust the topological extension direction of the fusion path and gradually generate multiple branch structure data fusion paths in a nonlinear manner. S336. The fusion order of each node in the real-time computing data fusion path structure is a nonlinear sorting function that combines the node's topological diffusion potential energy and topological centrality index. The hierarchical position of each node on the data fusion path is determined based on the fusion order. S337. Based on the determined data fusion path and fusion order, connect the topological relationships between data nodes level by level, establish a data fusion tree network with asymmetric topological branch structure point by point, and obtain a refined, nonlinear and multi-branch structure of cross-scale data fusion tree topology.

6. The energy consumption behavior simulation method based on a generative model according to claim 4, characterized in that, Specifically, S35 includes: S351. Taking each data point in the cross-scale fused data matrix as the center, determine the spatial neighborhood window centered on the data point. The size of the neighborhood window is a fixed spatial radius range centered on the data point. S352. Calculate the difference between the data point values ​​in the spatial neighborhood window and the center data point value one by one, and determine the degree of numerical deviation between the center data point and the spatial neighborhood data points based on the difference. S353. Based on the degree of numerical deviation of all data points in the neighborhood window, determine the local stability index of the central data point point by point. S354. Based on the local stability index of all data points in the cross-scale fused data matrix, calculate the stability index gradient of each data point in the spatial dimension. The stability index gradient is defined as the spatial weighted average of the difference between the local stability index of each data point and the local stability index of its neighboring data points. S355. Based on the stability index gradient of the data points, determine the spatial distribution intensity of the random noise weights point by point. S356. Based on the spatial distribution intensity of the calculated random noise weight, generate random noise values ​​point by point. The random noise values ​​follow a normal distribution with the random noise weight as the standard deviation and the mean as zero. S357. Based on the cross-scale fusion data matrix, the generated random noise values ​​are superimposed point by point onto the corresponding data points to finally generate a random noise weight matrix with a non-uniform spatial distribution.

7. The energy consumption behavior simulation method based on a generative model according to claim 1, characterized in that, S4 specifically includes: S41. Based on the reconstructed initial spatiotemporal energy consumption data matrix, calculate the nonlinear combination of spatiotemporal distance and numerical difference between each data point and other data points in the matrix point by point, so as to determine the characteristic spatial potential energy value of each data point. S42. Based on the characteristic spatial potential energy value of each data point, determine the potential energy spatial gradient vector of each position in the data matrix one by one through the spatial position coordinates. The magnitude and direction of the spatial gradient vector are clearly determined by the spatial partial derivative of the characteristic spatial potential energy value. S43. Based on the potential energy spatial gradient vector at each position in the data matrix, calculate and determine the spatial derivative of the potential energy gradient vector point by point to generate the spatial rate of change matrix of the potential energy gradient vector. S44. Based on the potential energy gradient vector space rate of change matrix, calculate the characteristic space gravitational intensity at each position in the data matrix space. The gravitational intensity value is determined by the absolute value of the value at each position in the potential energy gradient vector space rate of change matrix. S45. Based on all the gravitational intensities of the characteristic space determined in the data matrix, construct a clear characteristic space gravitational field point by point. The field strength at each position in the characteristic space gravitational field is directly represented by the numerical value of the characteristic space gravitational intensity at the corresponding position. S46. Based on the gravitational intensity distribution characteristics within the characteristic space gravitational field, and taking the relative magnitude of the local gravitational intensity as the basis, determine several spatial gravitational center nodes in the gravitational field. The gravitational center node is defined as the location point where the local gravitational intensity is extremely large in the characteristic space gravitational field. S47. Based on the determined position of the spatial gravitational center node and the gravitational field strength, dynamically determine the spatial position distribution of the initial quantum chaotic African vulture group, and determine the specific spatial position coordinates of the initial quantum chaotic African vulture group by non-uniform random sampling according to the position distribution.

8. The energy consumption behavior simulation method based on a generative model according to claim 1, characterized in that, S5 specifically includes: S51. Based on the spatial position coordinates of each individual in the initial quantum chaotic African vulture group in the characteristic space gravitational field, calculate the spatiotemporal gravitational gradient vector between each vulture individual and all gravitational center nodes in the gravitational field, and form the global gravitational perception vector of the vulture individual. S52. Based on the global gravitational perception vector of an individual vulture, the local chaotic perturbation probability at the current position is calculated in real time for each individual vulture. The local chaotic perturbation probability is determined by the nonlinear weighted sum of the change rates of the gravitational gradient vector at each position in the gravitational field of the characteristic space. S53. Based on the local chaotic perturbation probability, dynamically determine the perturbation threshold of the position update direction for each vulture individual, and use the perturbation threshold to adjust the position update step size of the vulture individual along the gravitational gradient vector direction in real time. S54. Calculate the dynamic spatial topological distance between the current spatial position of each vulture individual and the positions of each gravitational center node in the characteristic spatial gravitational field. The dynamic spatial topological distance considers a combination of Euclidean distance of position coordinates and nonlinear function of gravitational intensity difference. S55. Based on the calculated dynamic spatial topological distance, the target gravitational node of each vulture individual is dynamically determined. The target gravitational node is adjusted in real time and defined as the gravitational center node with the smallest calculated value of the combination of topological distance and gravitational intensity. S56. Each vulture individual is updated with its spatial coordinates gradually through a perturbation threshold, based on a dynamically determined target gravity node. The update method is constrained in real time by the combined effect of the gravity gradient vector and the probability of local chaotic perturbation. S57. Record the updated spatial coordinates of each vulture individual in real time, form a trajectory sequence for dynamic adjustment of the group position one by one, and finally obtain the vulture group position distribution trajectory sequence that dynamically changes with the characteristic spatial gravitational field, and determine the position distribution trajectory sequence as the dynamic optimization sequence of model parameters.

9. The energy consumption behavior simulation method based on a generative model according to claim 8, characterized in that, Specifically, S55 includes: S551. For each vulture individual's current position, construct multiple spatiotemporal paths from the vulture individual to each gravitational center node in the characteristic spatial gravitational field, and calculate the cumulative value of the spatial gravitational gradient of all data points on each path. The cumulative value of the spatial gravitational gradient is the point-by-point summation of the absolute values ​​of the gravitational gradient at each data point on the path. S552. Based on the cumulative value of the spatial gravitational gradient of each spatiotemporal path, calculate and determine the path space curvature index for each vulture individual. The path space curvature index is obtained by nonlinearly combining the cumulative value of the spatial gravitational gradient with the path length. The path length is defined as the number of data points on the spatiotemporal path. S553. Based on the spatial curvature index of the path from each vulture individual to each gravity center node, calculate the topological coupling strength between the paths one by one. The topological coupling strength is defined as a nonlinear function of the difference in spatial curvature index between each path, and generate a path topological coupling matrix centered on the vulture individual. S554. Based on the path topology coupling matrix, determine the global gravitational stability index of each path for each vulture individual. The global gravitational stability index is defined as the nonlinear weighted combination of the path space curvature index and the topology coupling strength. S555. Based on the global gravitational stability index of the path, sort the paths from each vulture individual to all gravitational center nodes, and determine in real time the current target gravitational node of the vulture individual as the gravitational center node connected by the path with the largest global gravitational stability index. S556. Monitor and record the changes in the current position coordinates and gravitational intensity of the characteristic space gravitational field of individual vultures in real time. When the position coordinates of individual vultures or the gravitational intensity in the characteristic space gravitational field change, recalculate and dynamically adjust the target gravitational nodes in real time to ensure that each individual vulture is always associated with the gravitational center node with the largest global gravitational stability index.

10. The energy consumption behavior simulation method based on a generative model according to claim 1, characterized in that, S6 specifically includes: S61. Based on the dynamic optimization sequence of model parameters, determine the range of cross-scale data overlap area corresponding to each optimization time in the sequence. The range of cross-scale data overlap area is determined based on the spatial location distribution trajectory sequence in the dynamic optimization sequence. S62. Based on the position coordinates of each data point in the cross-scale data overlap area, calculate the back-diffusion influence factor of each data point in the back-diffusion reconstruction process. The back-diffusion influence factor is defined as a nonlinear combination of the position distribution probability of the data point in the optimization sequence and the rate of change of the numerical value in the cross-scale fused data matrix. S63. Using the calculated inverse diffusion influence factor, dynamically adjust the data fusion weights during cross-scale data overlap and fusion point by point. S64. Calculate the dynamic decrease rate of random noise weights for each data point in the cross-scale fused data matrix based on the dynamic optimization sequence of the model parameters. S65. Based on the dynamically adjusted data fusion weight and the dynamic reduction rate of random noise weight, the cross-scale data overlap fusion and reverse diffusion process is executed point by point. The reverse diffusion process reconstructs the data scale by scale by gradually reducing random noise and tracing back the forward diffusion state. S66. Record the data reconstruction results in the reverse diffusion process in real time at each scale, and merge the recorded results at each scale to obtain the dynamic optimization spatiotemporal energy consumption behavior characteristic sequence after multi-scale reverse diffusion reconstruction, which is updated in real time with the dynamic optimization sequence of model parameters.

Images