Power distribution network source load space-time distribution collaborative prediction system based on deep learning model

Abstract

The invention discloses a power distribution network source load space-time distribution collaborative prediction system based on a deep learning model, and particularly relates to the technical field of power system operation and power distribution automation. The system accesses and aligns multi-source data under a unified time base, constructs graph priori carrying version weight in combination with operation topology, and outputs multi-step prediction of each node. Dynamic updating is introduced, wherein switching, reconstruction and power flow information is obtained through SCADA / power distribution automation, and graph priori is updated by event driving according to planned / instantaneous dual modes; performing time matching of the operation and the work order / instruction and historical plan probability weighting to obtain a planning coefficient, and setting a hysteresis threshold for discrimination; events such as tripping and protection action are directly judged to be instantaneous; and the undefined but continuously existing persons are conservatively classified according to the threshold duration. And the state duration threshold is self-adaptive along with the planning coefficient and the topological disturbance intensity. The system is matched with consistency auditing and gray release, and stable operation is supported.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system operation and power distribution automation, more particularly, the present application relates to a power distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model. BACKGROUND

[0002] The power distribution network source-load is strongly coupled in time and space: after the access of distributed photovoltaic, wind power, energy storage and electric vehicles, the fluctuation amplitude and uncertainty of node power increase significantly; within the same feeder, the power changes will propagate along the network due to the influence of line parameters and switch state between transformer areas. The available data sources in engineering are complex, but the time stamp caliber is different, there are delays and missing, and the sampling step is mixed, which often introduces bias when directly used for modeling. The existing methods are mainly based on single-point time series or simple spatial weighting of fixed topology, which is difficult to handle the graph structure changes caused by topology switching, planned maintenance / fault isolation, and also less explicit constraints and checking of feeder energy constraints and dimensionless. After the model goes online, there is a lack of participation of data quality labels and concept drift monitoring, and the versioning and gray release are incomplete, which leads to performance degradation and backtracking difficulty in real scenarios. SUMMARY

[0003] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a power distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model to solve the problems raised in the background art.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical scheme: The power distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model comprises the following modules: A data preprocessing module is used for accessing and aligning multi-source data from the power distribution operation side and the exogenous environment side under a unified time base, performing quality control, noise suppression and feature standardization on the multi-source data, and generating data quality labels aligned with the main data for each node at each time; A topology modeling module is used for constructing a graph structure prior based on the physical connection relationship and operating state of the power distribution network, and dynamically updating the graph structure prior through an event-driven dual-mode update mechanism; wherein, the planning coefficient is determined according to the time matching score and the historical planning probability, and the hysteresis threshold is set to distinguish between planned operations or instantaneous events; the planned operation is updated smoothly, and the instantaneous event is hard-switched in the next sampling period; the graph structure prior is normalized and bound with a topology version number of the time grid; A feature fusion module is used for constructing historical spatiotemporal fusion features, known exogenous features in the prediction period and spatial propagation operators based on the output of the data preprocessing module and the graph structure prior of the topology modeling module; a prediction model module, configured to input the historical spatiotemporal fused features, the predicted period future known exogenous features and the spatial propagation operator into a spatiotemporal deep learning model to perform source-load joint prediction, and output the predicted values of the load power and source power of each node in the future; a model management module, configured to train, deploy and perform online inference on the spatiotemporal deep learning model, and perform concept drift detection and model updating based on the running error and input distribution change.

[0005] In a preferred embodiment, the multi-source data includes load data from smart meters, switch state and power flow data from SCADA systems, output data of distributed power source monitoring terminals, meteorological elements from meteorological service platforms, and time labels from enterprise business management systems; the quality control includes backfilling of late data within a specified waiting window, entry of data beyond the window into a repair channel, and repair of abnormal and missing data by local time series interpolation and spatial neighborhood collaborative estimation.

[0006] In a preferred embodiment, the data quality label records whether the data is original reporting, late backfilling, missing interpolation, abnormal repair, and corresponding repair strategy and confidence level.

[0007] In a preferred embodiment, the edge weight of the graph structure prior is obtained by fusing a physical weight based on electrical parameters, a geographical weight based on geographical distance, and a statistical weight based on historical sequence covariation relationship.

[0008] In a preferred embodiment, the calculation method of the time matching score is to take the minimum time difference between the current operation time and the start / end time of the nearest planned maintenance work order or dispatching instruction, and calculate it based on the time difference and a tolerance time window.

[0009] In a preferred embodiment, in the topology modeling module, for a topology change that does not meet the planned operation and transient event criteria, when the state lasts more than a state duration threshold, it is determined as a transient event; the state duration threshold is adaptively adjusted according to the planned coefficient and the topology disturbance intensity coefficient.

[0010] In a preferred embodiment, the topology disturbance intensity coefficient is, in a unified evaluation range, the proportion of the number of affected connections to the total number of connections, and the proportion of the sum of the rated capacities of the affected branches to the sum of the rated capacities of all branches, and the larger of the two is taken as its value.

[0011] In a preferred embodiment, the feature fusion module enhances the model's ability to characterize behaviors at different time scales by constructing multi-scale time statistical features for key physical quantities, including local mean, fluctuation amplitude, quantile features, or typical profile similarity calculated within a fixed history window.

[0012] In a preferred embodiment, in the model management module, concept drift detection is based on rolling error indicators within a sliding window or divergence measures of input feature distributions from reference distributions during training; the model update mechanism includes fine-tuning for mild drift and retraining for severe drift.

[0013] In a preferred embodiment, in the model update mechanism, the switching of new and old models follows the process of shadow testing, gray release, and formal switching.

[0014] Technical effects and advantages of the present application: The present application processes event signals from protection devices, fault indicators, and intelligent switches in a unified time base, and unifies the processing of switch states, feeder reconstruction, and dominant flow indicators of SCADA and distribution automation, dynamically updates the graph structure prior according to the event-driven double mode (planning / instantaneous): when strong events such as fault tripping, protection action, and accident opening are detected, the instantaneous mode is directly entered to hard-switch the topology; in the non-strong event scenario, a planning coefficient is constructed by linearly fusing the time matching score and the historical planning probability, and an entry / exit hysteresis threshold is configured to robustly determine the planning or instantaneous mode; for the transient state in the determination interval, the last determination is maintained to suppress frequent jitter. For the topology changes that do not meet both criteria and continue to exist, a state duration threshold determined by the planning coefficient and the topology disturbance intensity coefficient is set as a bottom line, and if the duration exceeds the threshold, the instantaneous mode is entered according to the conservative strategy and an alarm is generated. Among them, the topology disturbance intensity coefficient takes the evaluation sub-network / feeder as the range, and takes the larger one of the affected connection scale proportion and the affected branch capacity proportion, considering the change scale and capacity weight. The above mechanism ensures that the topology prior can respond in time without jitter in situations such as tie-in / out, distributed power on / off, and flow direction change, weakens the interference of false judgment and false switching on space propagation, and realizes the balance of responsiveness and stability.

[0015] On the model side, the prior map is normalized and bound with the time grid version number, serving as a spatial propagation operator to participate in the source-load joint multi-step prediction of spatio-temporal deep learning; quality control and standardization of multi-source data ensure consistent and traceable caliber during training and online inference. Thus, the model uses the topology and exogenous information matching the current running state at each time, avoiding propagation distortion and error steep rise caused by topology drift; the hysteresis + bottom-up strategy reduces false triggering and alarm noise under abnormal working conditions; based on the association of versioned prior and event log, it is convenient for playback, verification and operation and maintenance closed loop. The overall effect is reflected in: the adaptability to frequent reconstruction and strong disturbance scene is improved, the stability and consistency of the prediction result are enhanced, the maintainability and auditability of the whole life cycle are enhanced, so as to obtain more reliable source-load collaborative prediction output under the real power distribution network operation condition. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to facilitate the understanding of those skilled in the art, the present application will be further described below in conjunction with the drawings; Figure 1 Fig. 1 is a structural schematic diagram of the power distribution network source-load spatio-temporal distribution collaborative prediction system based on the deep learning model of the present application; Figure 2 Fig. 2 is a flowchart of the topology change event discrimination and dynamic updating of the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] Embodiment 1, the power distribution network source-load spatio-temporal distribution collaborative prediction system based on the deep learning model of the present application, as shown in Figure 1 includes the following modules: Module 1: data preprocessing module The purpose of this module is to build a high-quality data foundation that can directly support spatio-temporal collaborative modeling. The system establishes a stable data access channel with existing collection and monitoring systems in the power distribution network operating environment, and continuously acquires observation data and exogenous data from AMI / smart meters, distribution automation / SCADA, distributed power monitoring terminals, environmental and meteorological service platforms, and enterprise business management systems, etc. The data types involved include but are not limited to: historical active / reactive load power at the substation or feeder level, distributed power output, node voltage and current, switch and tie-in state, device operation alarm and maintenance mark, measurement point geographic information and asset coding, meteorological elements, and time labels, etc. Distributed power includes photovoltaic, wind power, etc., meteorological elements include temperature, humidity, wind speed, wind direction, irradiance, rainfall, etc., and time labels include year / month / day / hour, weekday / weekend, holiday and special event day, etc. Link encryption and access control strategies are used during data access and transmission, and sensitive fields are desensitized and permission isolated before entering the modeling domain to achieve information security and compliance.

[0019] To ensure that cross-source data has consistent standards and comparability, the system first establishes a unified time reference and target sampling step, aligns data from different sources and different sampling frequencies to a unified time grid through clock synchronization and timestamp correction; for power-type continuous quantities, interval averaging or equivalent mapping is implemented according to the target step to aggregate high-frequency samples into representative values of the step; for energy-type quantities represented by reading increments, the difference between adjacent period readings is taken and converted into the equivalent average power of the period, thereby eliminating statistical bias introduced by measurement caliber differences. In view of the objective existence of field communication jitter and late reporting, the system configures a late buffer mechanism: late samples that arrive within the specified waiting window are backfilled according to their original time grid; samples that exceed the window are entered into the repair channel as supplementary data and record the source and time lag information in the quality label to support subsequent training weighting and full-process tracing.

[0020] After time alignment, systematic data quality control is carried out. First, structural and scope consistency checks are performed, including unit and dimension consistency, device code and geographic coordinate validity checks, rationality judgments of measurement upper and lower limits and engineering thresholds, and coarse checks of balance relationships and physical constraints. Then, time consistency and dynamic reasonableness tests are performed. Suspicious samples are identified through three types of signal detection: long-time constant values (flat lines), sudden step amplitude over-limit, and too fast change rate. Cross-validation is performed in combination with device alarms and maintenance windows. Finally, cross-source consistency audits are performed, such as aggregating node-level net load to the feeder end and reconciling with feeder measurements. When significant deviations are found, they are marked as structural abnormalities and enter the special repair strategy. For abnormalities and missing values marked by rules and statistical detection, the system performs interpolation and repair according to the following hierarchical order: local time sequence first, neighborhood space assistance, and exogenous information correction. Specifically: in the case of short continuous missing, local time interpolation or limited forward filling is used to maintain trend continuity; in the case of longer gaps or structural abnormalities, the related node curves of the spatial neighborhood are introduced for collaborative estimation, and meteorological elements and typical day patterns at the same time are used for consistency correction; when there are device replacement or topology switching events, event-based segmented modeling is preferred to avoid statistical pollution across events. All repaired or recorded samples will record the type of missing / abnormal, repair strategy and confidence level in the quality label, and the version information of the original and repaired values will be preserved to support audit and backtracking.

[0021] To improve the signal-to-noise ratio while avoiding weakening key dynamic characteristics, the system implements mild noise suppression after removing outliers. Preferably, short window median filtering or low-order polynomial smoothing is used, and for high-volatility exogenous quantities such as photovoltaic irradiance and wind speed, an adaptive window width is used to balance the sensitivity to sudden changes in sunlight and the suppression of random jitter. The smoothing parameter is determined based on the historical validation set, and small adjustments are made through performance monitoring after going online to ensure fidelity to short-term peaks and valleys such as afternoon photovoltaic sudden drop and evening peak rapid rise.

[0022] After noise processing, the system unifies the scale and coding method in all feature dimensions. Numerical features are standardized or interval scaled based on training set statistics, and the scaling parameters are frozen to prevent target leakage; periodic time labels are represented by embedding or sine-cosine position encoding to represent their seasonality and intra- / inter-week periodicity; categorical labels are converted into sparse embeddings or one-hot encodings for feature sets. For exogenous quantities from meteorological grid points or regional stations, the system maps them to the node domain based on the spatial relationship between nodes and grid points. The mapping weights are determined by geographic distance, terrain / shading information, or historical correlation, and are guaranteed to be non-negative and normalized to obtain the exogenous driving sequence for each node at the unified time grid.

[0023] To enable subsequent training to distinguish sample credibility, the system generates a data quality label aligned with the main data for each node at each time, which includes at least the following fields: whether it is originally reported, whether it is late backfilled, whether it is missing interpolated, whether it is abnormal repaired, whether it is passed through cross-source checking, corresponding repair strategy and credibility classification, etc. The label neither changes the feature value itself nor performs weighted calculation in this module, but is used as the basis for loss weighting and sample screening in the training stage, and is used as the basis for result confidence and explainability prompt in the inference stage. Combined with the quality label, the system also precipitates meta-information of data version, feature dictionary and scaling parameter, and binds it with timestamp and topology version number to ensure data processing consistency and traceability in the training, validation and deployment link.

[0024] Through the above data collection and unified processing, this module outputs a data foundation with time consistency, dimension consistency, source traceability and quality measurability, providing reliable, sufficient and engineered input conditions for subsequent power grid topology graph modeling, feature construction and spatio-temporal collaborative prediction. On this data foundation, the system can stably support modeling needs in different regions, different voltage levels and different new energy penetration scenarios, thereby ensuring the implementability and long-term maintainability of the overall invention in real distribution network environment.

[0025] Module 2: Topology modeling module The purpose of this module is to abstract the physical connection relationship, operating state and historical statistical characteristics of the distribution network into a graph structure prior that can be directly used by the spatio-temporal deep model, so that the system can accurately express the coupling and influence path between nodes in the spatial dimension and adaptively update with the evolution of topology and power flow during operation. The system takes a unified time grid as the alignment benchmark and constructs a graph model under the condition of consistent data time index with the output of module 1. For ease of description, the distribution network is abstracted as a graph structure G=(V,E,A), where V is the set of nodes participating in modeling, typically corresponding to substation, distribution transformer, feeder end or key measurement point accessing distributed power source; E is the edge set, representing the electrical accessibility relationship between nodes through feeders, transformers or tie switches; is a weighted adjacency matrix, and the element characterizes the node and the spatial correlation strength under the current modeling perspective, indicates that they are not adjacent under this perspective. The nodes and edges of the graph structure have a topology version number associated with the time grid to ensure that the subsequent model uses a matching spatial prior at any time step.

[0026] The determination of edge weights considers three aspects of information, i.e., electrical mechanism, geographical proximity and historical statistical correlation, to balance the physical interpretability and data-driven adaptability. The electrical mechanism is derived from the line parameters and equipment capacity limits, such as equivalent line impedance, rated current / capacity, typical power flow distribution, etc., to characterize the difficulty of energy and disturbance transmission in the network; the geographical proximity is based on GIS coordinates, line path and shielding conditions, to reflect the spatial consistency of external environmental disturbances; the historical statistical correlation is given by the correlation coefficient or mutual information of historical source / load time series between nodes, to reveal the covariation relationship that cannot be directly explained by static parameters. To fuse the above three types of information into a single edge weight, the system adopts a linear weighting method to obtain the comprehensive weight, as follows: , where is the physical weight based on electrical parameters and equipment capacity constraints, is the geographical weight based on geographical distance and path conditions, is the statistical weight based on historical sequence covariation relationship; , , are fusion coefficients, , , ≥ 0 and + + = 1. To avoid over-amplification of weak links in model propagation, the system applies scaling to according to the capacity limit of the equipment corresponding to the edge, so that the edge weight close to or below the measurable noise level is numerically suppressed; for long-term shutdown, planned maintenance or isolated links, the edge weight is directly reduced to zero.

[0027] Considering the time-varying nature of distribution networks in actual operation, this module introduces a dynamic updating mechanism to reflect the impact of topology switching, tie switch switching, distributed power supply access / exit and power flow direction change on spatial priori. As shown in Figure 2 , the system obtains the switch state, feeder reconstruction information and dominant power flow direction indication at each time through SCADA and distribution automation system. To achieve accurate response, the system discriminates the topology change event according to multi-source information, and adopts an event-driven double-mode updating mechanism. The system obtains advanced event signals from protection devices, fault indicators and intelligent switches through standard data interface, and once one or more of fault tripping, protection action, accident opening are identified, it is immediately determined as mode two transient event. A planned coefficient is defined to unify the discrimination: first calculate the time matching score and the historical planning probability , where ;the minimum time difference between the current operation time and the start / end time of the nearest planned maintenance work order or dispatch instruction, = 15 min; if there are multiple candidate work orders / instructions, take the minimum value of Linearly fuse the two to obtain the planned coefficient , as follows: ; wherein and are weights, for example = 0.4, = 0.6.

[0028] The historical planning probability refers to the proportion of the same type of operation triggered by an effective work order or dispatch instruction in a historical rolling window under the condition of a given intraday time slice and day type, and the prior probability obtained by smoothing processing, with a value range of [0, 1].

[0029] Based on the coefficient, enter / exit hysteresis threshold values are set: when ≥ θ1, it is determined to be mode one planned operation, when ≤ θ2, it is determined to be mode two instantaneous switching, and the last determination is maintained in the middle interval to avoid frequent switching; for example, θ1= 0.75, θ2= 0.65.

[0030] For topology changes that do not meet the mode two and mode one criteria, when the state lasts more than the state duration threshold, the system adopts a conservative strategy to determine it as the default mode two instantaneous switching, and generates an alarm log and pushes it to the operator console.

[0031] The state duration threshold is determined by the planned coefficient and the topology disturbance intensity coefficient, as follows: ; is the lower limit of the duration, is the upper limit of the duration, for example = 1 min, = 10 min; is the topology disturbance intensity coefficient; the topology disturbance intensity coefficient is used to represent the relative influence intensity of a topology change in the evaluation subnetwork / feed line, and the larger the value, the stronger the disturbance. Its physical meaning covers both the scale and the capacity dimensions: under a unified time grid, first determine the connection set involved in this change and the target subnetwork / feed line to which it belongs as the evaluation range; then calculate the scale proportion and the capacity proportion, and take the larger proportion of the two as the value of the topology disturbance intensity coefficient. The scale proportion is the proportion of the number of affected connections to the total number of connections in the evaluation range, and the capacity proportion is the proportion of the sum of the affected branch rated capacities to the sum of the rated capacities of all branches in the evaluation range.

[0032] For the topology change determined as a planned operation, the system updates the adjacency matrix using an exponential smoothing transition strategy, with a smoothing time constant set to 3 sampling periods, and realizes the smooth transition of the topology version through weighted average. For transient events, the system completes the hard switching of the adjacency matrix in the next sampling period, directly uses the adjacency matrix version corresponding to the new topology, and skips the smooth transition link. The disconnected edges are removed from the adjacency matrix, and the edge weight of the conducting edge is restored to its fusion value; in the scenario where direction-sensitive relationships need to be expressed, the system further constructs a directed adjacency representation, so that the propagation from upstream to downstream and the reverse propagation are distinguished numerically.

[0033] To meet the requirements of numerical stability and convergence for subsequent graph neural network calculations, the system performs normalization processing on the adjacency matrix and introduces self-loops, so as to preserve node self-information, control the spectral radius and stabilize training during spatial propagation; the normalization method can be symmetric matrix normalization or other equivalent numerical processing methods in engineering, and the specific implementation is not limited. If it is necessary to enhance the perception ability of distant neighbors, the system prepares multiple scale adjacencies at the same time, so that the model can aggregate information at different spatial scales, thereby better expressing phenomena such as branch linkage, power flow migration and local congestion. All normalization and multi-scale representation are consistent with the time grid and are bound to the topology version number, ensuring that any time step uses a spatial propagation operator that matches it.

[0034] To ensure that the graph prior is consistent with the power physical constraints, this module includes a consistency check and abnormal repair process after construction. The system checks the connectivity of each feeder subgraph to avoid numerical anomalies caused by isolated nodes or hanging edges in spatial propagation; in scenarios that require radial operation according to regulations, the system performs loop detection and applies penalties or forced ring breaking to abnormal closed loops; at the energy convergence level, the system checks the node-level net load summary and feeder measurements, locates and traces abnormal sections that exceed the tolerance to the corresponding edge or node label. The above checks are not limited to specific implementation methods, as long as they can improve the physical consistency and explainability of the graph prior without changing the true operating state.

[0035] The system persistently stores the graph prior in a structured form of node list, edge list, weight field, topology version number, and time index, and incorporates the unified metadata management of the feature dictionary, scaling parameter, and data quality label generated by the module. The model calling end retrieves the corresponding adjacency representation through the time index, realizes consistent access to historical backtesting, online inference, and offline training, and automatically generates a new version of the graph prior and retains the historical version to support tracing and comparison when the topology changes during operation or the historical data is revised by supplementing the record. After processing by this module, the obtained graph structure prior can provide stable, accurate, and evolving spatial relationship support for subsequent spatial-temporal feature encoding and source-load collaborative prediction without increasing the running burden, thereby significantly improving the model's ability to describe the real physical relationship and spatial propagation law of the distribution network.

[0036] Module three: feature fusion module This module takes the uniformly processed data base output by module one as input and completes the integrated representation construction of each node of the distribution network in the time and space dimensions under the constraint of the power grid topology graph prior determined by module two. The system aggregates historical observations, future known exogenous quantities, and time semantic information on a unified time grid for each modeling node, and forms a high-dimensional structured representation that can be directly fed into the spatial-temporal collaborative prediction model under the premise of introducing spatial prior. To avoid interference of the same physical quantity in different dimensions on the training process, numerical features are standardized using the scale parameter frozen by module one; discrete time labels use embedded or equivalent periodic position encoding to retain seasonality and intraday / weekly periodicity, and discrete time labels include hours, weeks, weekdays / holidays, and special event days. For node operation state related category labels, such as device maintenance and switching event placeholders, they are incorporated into the feature set in a sparse embedding manner to ensure time consistency, so that the model can identify the disturbance effect of events on source-load behavior patterns during the learning process.

[0037] To make the exogenous driving correspond one-to-one with the electrical nodes in space, the system maps meteorological elements from meteorological stations or grid points to the node domain. Let the set of meteorological grid points be , and the rth element value of any grid point g at time step t be ; calculate the non-negative and sum-to-1 spatial mapping weight for each node i, which can be determined by integrating geographical distance, terrain obstruction, and historical correlation. On this basis, the exogenous elements of the node domain are aggregated by weight to obtain . The exogenous quantities known in the future within the prediction period are kept in the same time grid and aligned to the node domain to ensure that the decoding stage can be conditioned with the corresponding future time, such as hourly weather forecasts, peak electricity prices, or demand response calendars. The above mapping relationship does not change the physical meaning of the exogenous quantity, only completes the spatial reference frame conversion, and the weight calculation method is not limited to a specific algorithm, but must satisfy the non-negative and normalized constraints to ensure the interpretability of energy and statistics.

[0038] Considering that the power load and distributed power output have both short-term fluctuations and slow-changing trends, the system constructs multi-scale time statistical features for key physical quantities without introducing additional trainable parameters, to enhance the ability to describe different time scale behaviors. Specifically, the local mean, fluctuation amplitude, quantile feature, or intra-day / week similarity of the typical profile is calculated within a fixed historical window, and these statistics are aligned with the original standardized sequence in the time dimension and then spliced. This processing can provide the model with explicit description of background level and abnormal deviation without changing the original time sequence, improving the robustness in extreme weather, load migration, or user behavior mutation. To further improve spatial distinguishability, the system performs a linear neighborhood propagation with the normalized adjacency matrix constraint given by Module 2 before entering the deep model, to obtain a first-order spatial prior smoothing result; this propagation does not change the feature meaning, but only aggregates the neighborhood information with the topological prior as the weight, thereby providing a more stable numerical starting point for the subsequent graph propagation layer.

[0039] In terms of data quality, the quality label corresponding to the main data generated by Module 1 is kept aligned in this module and serves as accompanying information throughout the subsequent modeling process. The quality label does not change the feature value itself, but is mainly used for loss weighting and sample selection in the training stage, as well as result reliability prompt and explainability output in the inference stage. For samples generated by late backfilling, missing interpolation, or structural anomaly repair, the quality label will accurately record the source, strategy, and confidence level to reduce the impact of low-confidence samples on parameter estimation during model learning.

[0040] To facilitate the description of the output structure of this module, the historical spatiotemporal fusion feature tensor is defined as where is the number of nodes participating in modeling, is the length of the historical time window, is the dimension of the fused historical features; the future known exogenous feature tensor in the prediction period is defined as where is the number of prediction steps, is the dimension of the exogenous features. To match this, Module 2 forms a symmetric normalized adjacency matrix and its version over time) are provided as spatial propagation operators together with the output of this module. At this point, the system completes the fusion and structured packaging of the three types of information, i.e., historical observations, exogenous driving, and topological priors, in a unified spatio-temporal reference system, obtaining the input triplets of . The triplets not only maintain the high-resolution observable information at the node level, but also explicitly encode the neighborhood spatial coupling and cross-period exogenous constraints, providing a consistent, sufficient, and engineering-implementable input basis for the spatio-temporal collaborative prediction model of module four.

[0041] Module four: prediction model module Based on modules one, two, and three, this module constructs a spatio-temporal deep learning model for source-load joint prediction, enabling the system to simultaneously predict the future load power and distributed power output of each node in the distribution network for several time periods given historical observations and known future exogenous driving. The model takes the triple input output by module three as a precondition, i.e., the historical spatio-temporal fusion feature tensor, the future known exogenous feature tensor for the prediction period, and the symmetric normalized adjacency matrix aligned with the time grid. For ease of explanation, let the historical spatio-temporal fusion feature tensor be , with dimensions , where represents the number of nodes involved in modeling, represents the length of the historical time window, represents the dimension of the fused historical features; let the future known exogenous feature tensor for the prediction period be , with dimensions , where represents the number of prediction steps, represents the dimension of the exogenous features for the prediction period; let the symmetric normalized adjacency matrix be , with dimensions , representing the spatial dependency between nodes under the current topology version. The overall mapping relationship of the model can be summarized as , where is the spatio-temporal neural network determined by the parameter set θ, outputting with dimensions N × H × 2, where the last dimension 2 corresponds to the point prediction values of load power and source power, respectively. The above symbols are only used to define the meanings and dimensions of inputs and outputs, without limiting the specific network implementation form.

[0042] In terms of structure design, the model follows the overall idea of encoding and jointly decoding spatial, temporal, and exogenous driving in stages. First, for each time step within the historical time window, the model utilizes the The spatial encoding is performed on the node features at the same time, refining the coupling information in the local network in a neighborhood-aggregated manner to form a numerically stable and physically consistent spatial representation. This spatial representation is computed in parallel on all nodes, preserving the running portrait of each node while explicitly absorbing the state changes of its directly or indirectly neighboring nodes, thus providing a spatially annotated historical trajectory for subsequent time modeling. Second, the model encodes the historical spatial representation sequence of each node in time, learning the temporal evolution rules through a structure that can capture both short-term fluctuations and long-term dependencies. In this process, the model automatically identifies key moments that have a more significant impact on the future, such as high-temperature periods, sudden changes in afternoon radiation, or evening rush hours on weekdays; and gives higher weights in the encoding results, so that the historical behavior of the node level is compressed into discriminative time representation vectors in the form of context awareness.

[0043] To ensure the temporal consistency of the prediction and the exogenous driving outside the prediction period, the model introduces a conditioning mechanism for future time steps in the decoding stage: for each prediction time step, the model aligns the historical time representation with the exogenous features corresponding to that time step as conditional information, selects the most relevant historical patterns for that exogenous scenario, and fuses them to obtain a comprehensive representation consistent with the semantics of that future time. The exogenous features corresponding to that time step, such as hourly weather forecasts, peak electricity prices, or demand response calendar placeholders. This process does not change the physical meaning of the exogenous quantities, but aims to allow the model to see the known external driving forces at different future time steps when generating predictions, and to retrieve and reweight the historical trajectory accordingly. After spatio-temporal fusion, the model forms a unified prediction feature vector at each node and each future time; then, through a multi-task regression head, it simultaneously outputs point predictions of load power and source power. For business interfaces that need to provide net load externally, the system calculates the net load in a derived relationship, i.e., subtracts the source power prediction from the load power prediction at the same node and future time. The derived operation is completed outside the main network structure, without changing the structure of the main network.

[0044] This module performs time alignment on the adjacent version of in the spatial encoding link: both the historical time step and the spatial propagation of the prediction period use the adjacent version consistent with its timestamp to reflect the immediate impact of switch commissioning, feeder reconstruction, or distributed power supply access / exit on the coupling between nodes. To avoid numerical jitter caused by frequent switching, the system uses a smooth transition strategy for adjacent version updates, and speeds up the transition when there is a clear topology switching event, so that the spatial encoding is sensitive enough to topology changes and maintains the stability of the reasoning process. In conjunction with this, the time encoding and decoding part imposes consistency constraints on the length, step size, and mask strategy of the input sequence, ensuring that the model can still handle missing or delayed backfill samples robustly based on quality markers.

[0045] Module five: model management module On the basis of the spatio-temporal collaborative prediction model structure described in module four, this module completes the training sample construction and parameter optimization, online inference service deployment, and concept drift dynamic updating mechanism for the running period, thereby ensuring that the power distribution network source and load spatio-temporal distribution collaborative prediction system based on the deep learning model maintains stable, interpretable, and evolving prediction performance in the long-term, complex, and changing power grid environment. The system takes the historical spatio-temporal fusion feature tensor output by module three, the future known exogenous feature tensor in the prediction period, and the symmetric normalized adjacency matrix aligned with the time grid as the model input, and combines the unified time grid and quality label to complete the alignment of the supervision signal and the generation of the sample. For convenience of definition, let the historical spatio-temporal fusion feature tensor be ; let the future known exogenous feature tensor in the prediction period be ; and let the symmetric normalized adjacency matrix be . The overall mapping relationship of the model is still denoted as The above formula is only used to define the meaning and dimension of the input and output, and does not limit the specific network implementation form.

[0046] In the training phase, the system uses a sliding time window strategy to construct samples on the unified time grid to ensure time sequence continuity and scene coverage; the supervision signal is the true load power and source power of each node in the future H time steps, and the quality label is aligned with the sample one by one for loss weighting and sample screening, thereby reducing the interference of missing interpolation, abnormal repair, or late backfill samples on parameter learning. Parameter optimization takes a multi-objective joint criterion as the core, focusing on both point prediction error and spatial consistency and physical consistency as well as optional uncertainty expression. The point prediction error is quality-weighted summed in the node-time two-dimensional space through the average absolute error or the mean square error, highlighting the learning contribution of high-reliability samples; the spatial consistency is inhibited through the smoothing penalty based on the graph Laplacian to suppress the mutation of adjacent nodes in the absence of physical basis, making the spatial propagation more stable in value and more reasonable in physics; the physical consistency is audited and corrected through the energy convergence constraint at the feeder level, ensuring that there is no systematic deviation between the node-level prediction and the upper-layer measurement.

[0047] For ease of understanding, the feeder energy consistency can be expressed as , where represents the measured active power injection of the feeder in the future step , ∈{0, 1} is the node and feeder attribution matrix element, and takes 1 to represent that the node i belongs to the feeder f, is the net load prediction of node i at future step h. The above relationship is incorporated into the joint objective in the form of a penalty term or constraint during training to improve engineering interpretability and alignment with upper-level reconciliation. In cases where interval prediction or confidence output is required by the business, a quantile channel is added at the decoding end of the model to achieve uncertainty characterization through quantile loss; if only point prediction is needed, the relevant channel can be closed as needed without affecting the main training process. The optimization algorithm uses a small batch iterative adaptive gradient method, combined with learning rate scheduling and gradient clipping to ensure numerical stability; the validation set is divided in a time-forward rolling manner, and the early stopping criterion is determined according to the improvement amplitude and consecutive rounds of the joint index, so as to balance convergence and generalization ability.

[0048] After completing the training and achieving the online standard through historical backtesting and offline testing, the model is deployed in the inference environment in a service-oriented form. The deployment side solidifies the feature dictionary, scale parameters and adjacent version snapshots to ensure consistency with the training side; the exposed inference interface takes the input triple consistent with the module four as input, and returns the node-level source load joint prediction results and optional uncertainty information for the next H time steps. The system continuously records key performance indicators including the average absolute percentage error, root mean square error, interval coverage rate and interval width aggregated by nodes and feeders during the online phase, and compares the distribution during the running period with the benchmark distribution during the training period to form a visual performance baseline and fluctuation band. For known events such as topology switching, extreme weather or abnormal user behavior, the system aligns with the input triple through the topology version number bound to the time grid and the event placeholder on the calling side, ensuring consistent decoding with the actual situation at the time.

[0049] The concept drift detection and update mechanism is configured based on error monitoring and statistical test. The system calculates a rolling error indicator in a sliding window, and determines that concept drift occurs when the indicator continuously exceeds a threshold value or when a divergence measure between the input feature distribution and the reference distribution during the training period significantly increases. When mild drift occurs, the system preferentially performs a fine-tuning process: using new data in the near neighbor period as a sample, the decoding end and the output end are retrained with small steps, and the spatial coding layer is usually fixed or semi-fixed to reduce the update cost and avoid overfitting; when severe or continuous drift occurs, the system enters a retraining process: the training set is reconstructed with the latest monthly or quarterly data, and the full model training and backtest evaluation are completed again. The switching of new and old models follows the engineering strategy of shadow testing, gray release and formal switching: the new model first reasons in the shadow channel in parallel with the existing network model, compares key indicators and physical consistency check items, and after meeting the threshold, a small amount of traffic is diverted for gray release, and finally the full switching is completed after the stable period is met; if abnormal fluctuations or indicator degradation occur in the short term, it is automatically rolled back to the previous stable version. Each update is registered in the model repository, and the registration content at least includes the training data time range, the main hyperparameters and loss weights, the key evaluation indicators and the online approval records to meet the audit and traceability requirements.

[0050] Under the above-mentioned training, deployment and continuous updating mechanism, the system realizes comprehensive protection of point prediction accuracy, spatial smoothing consistency, feeder energy constraint consistency and uncertainty expression; in the running period, fine-tuning or retraining is triggered by error and distribution double threshold detection, and strict version management and gray release strategy is used to maintain the stability and controllability of online service. Therefore, the present application can maintain high reliable and evolvable source and load spatio-temporal distribution collaborative prediction ability in real power distribution network scenarios with multiple regions, multiple voltage levels and multiple new energy penetration rates for a long time, and meet the comprehensive requirements of engineering application on accuracy, interpretability and life cycle management.

[0051] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0052] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the embodiments of the device described above are merely schematic, and the division of the units is merely logical function division. There can be other division manners in actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0053] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments of the present application.

[0054] In addition, each functional unit in the embodiments of the present application can be integrated in a processing unit, or each unit can exist physically as a separate unit, or two or more units can be integrated in one unit.

[0055] The above describes only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A collaborative prediction system for the spatiotemporal distribution of power distribution networks based on a deep learning model, characterized in that, Includes the following modules: The data preprocessing module is used to access and align multi-source data from the power distribution operation side and the exogenous environment side under a unified time base, perform quality control, noise suppression and feature standardization on the multi-source data, and generate data quality labels aligned with the master data for each node at each time. The topology modeling module is used to construct graph structure priors based on the physical connections and operating status of the distribution network, and dynamically update the graph structure priors through an event-driven dual-mode update mechanism. Specifically, the planning coefficient is determined based on the time matching score and historical planning probability, and a hysteresis threshold is set to distinguish between planned operations and transient events. Planned operations are updated smoothly, while transient events are hard-switched in the next sampling period. The graph structure priors are normalized and bound to the topology version number with the time grid. The feature fusion module is used to construct historical spatiotemporal fusion features, known exogenous features for the future prediction period, and spatial propagation operators based on the output of the data preprocessing module and the graph structure prior of the topology modeling module. The prediction model module is used to take historical spatiotemporal fusion features, known exogenous features in the prediction period and spatial propagation operators as inputs, and perform joint source-load prediction through spatiotemporal deep learning model to output the predicted load power and source power of each node in the future multiple steps. The model management module is used for training, deploying, and inferring spatiotemporal deep learning models, and performs concept drift detection and model updates based on runtime errors and changes in input distribution.

2. The distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model according to claim 1, characterized in that: Multi-source data includes load data from smart meters, switch status and power flow data from SCADA systems, output data from distributed power monitoring terminals, meteorological elements from meteorological service platforms, and time stamps from enterprise business management systems. Quality control includes backfilling late data within a specified waiting window, entering supplementary data that exceeds the window into a repair channel, and repairing abnormal and missing data using local temporal interpolation and spatial neighborhood co-estimation.

3. The distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model according to claim 1, characterized in that: Data quality labels record whether the data is the original report, whether it is a late backfill, whether it is missing data imputation, whether it is abnormal data repair, and the corresponding repair strategy and trust level.

4. The distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model according to claim 1, characterized in that: The prior edge weights of the graph structure are obtained by fusing physical weights based on electrical parameters, geographical weights based on geographical distance, and statistical weights based on historical sequence covariance relationships.

5. The distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model according to claim 1, characterized in that: The time matching score is calculated as follows: take the minimum time difference between the current operation time and the start / end time of the most recent planned maintenance work order or scheduling instruction, and calculate it based on this time difference and a tolerance time window.

6. The distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model according to claim 1, characterized in that: In the topology modeling module, for topology changes that do not meet the criteria for planned operations and transient events, if the state continues for more than the state duration threshold, it is determined to be a transient event; the state duration threshold is adaptively adjusted according to the planning coefficient and the topology disturbance intensity coefficient.

7. The distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model according to claim 6, characterized in that: The topology disturbance intensity coefficient is calculated within a unified evaluation range, taking the larger of the two ratios as its value: the ratio of the number of connections affected by this topology change to the total number of connections, and the ratio of the sum of the rated capacities of the affected branches to the sum of the rated capacities of all branches.

8. The distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model according to claim 1, characterized in that: The feature fusion module enhances the model’s ability to characterize behavior at different time scales by constructing multi-scale time statistical features for key physical quantities. These multi-scale time statistical features include local mean, fluctuation amplitude, quantile features, or typical contour similarity calculated within a fixed historical window.

9. The distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model according to claim 1, characterized in that: In the model management module, concept drift detection is based on the rolling error index within the sliding window or the divergence measure between the input feature distribution and the training period reference distribution; the model update mechanism includes fine-tuning the model with slight drift and retraining the model with severe drift.

10. The distribution network source-load spatiotemporal distribution collaborative prediction system based on a deep learning model according to claim 1, characterized in that: In the model update mechanism, the switching between old and new models follows the process of shadow testing, canary release, and official switching.

Images