Office park electric load forecasting and dispatching method for novel electric power system

Abstract

The invention belongs to the technical field of office park electric load prediction, and relates to an office park electric load prediction and scheduling method for a novel electric power system. The method comprises the steps that S1, multi-source data fusion and panoramic view construction are carried out, historical load data of a park are collected, and park operation data, meteorological data and a special event calendar are obtained; s2, data preprocessing and feature directional extraction; s3, building and cooperating short-term, medium-term and long-term prediction models on the basis of multi-time scale prediction of model cooperation to form a hierarchical prediction system; s4, optimizing and evaluating a prediction result, and performing uncertainty evaluation on the prediction result in the step S3; and S5, generating and executing a hierarchical scheduling instruction based on hierarchical scheduling decision and execution of multi-objective optimization. According to the method, the problem of full-chain link splitting is solved, panoramic data view, collaborative prediction model, quantitative risk assessment and multi-target optimization scheduling are realized, prediction accuracy and scheduling flexibility are improved, and active participation in energy management of the office park is supported.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of office park electric load prediction, and particularly relates to an office park electric load prediction and dispatching method for a new power system. BACKGROUND

[0002] In the field of office park energy management, the traditional electric load prediction and dispatching method has long been facing the challenge of fragmentation of the whole chain. The data collection link only relies on historical load records, and fails to effectively integrate multi-source heterogeneous information such as operating status, weather conditions and special events, resulting in scattered data views and insufficient spatio-temporal consistency, which is difficult to support panoramic energy analysis. In the prediction link, short-term, medium-term and long-term load prediction models operate independently, lack of internal coordination mechanism, and cannot realize the organic integration of time sequence dynamic characteristics and structured static characteristics, resulting in disconnection of prediction results in time scale, especially difficult to capture the correlation influence of calendar effect and trend change. The risk assessment link is often simplified or completely missing, resulting in that the dispatching decision cannot quantify the risk of extreme scenarios, and the strategy is only based on deterministic prediction results, which is difficult to adapt to the fluctuation of renewable energy and the demand of power grid interaction in the new power system. The dispatching link is limited to single economic target optimization, ignoring the multi-objective balance of local renewable energy consumption and power supply reliability, so that the office park is always in a passive power receiving state and cannot actively participate in demand response or dynamically adjust distributed resources.

[0003] With the accelerated construction of the new power system mainly based on new energy, a large number of office park users need to transform from rigid load points to intelligent energy nodes with active adjustment ability, and realize two-way flexible interaction with the power grid. However, the existing technical solutions focus on local improvement of isolated links, and fail to build a closed-loop system of data utilization, model construction, risk assessment and intelligent decision-making, lacking systematic integration of multi-source data fusion, model collaborative optimization and uncertainty quantification, which seriously restricts the resource coordination ability and operation flexibility of the park in the new power system. Therefore, an office park electric load prediction and dispatching method for a new power system is needed to solve the problem of link fragmentation and promote the transformation of office parks to active energy management paradigm. SUMMARY

[0004] Therefore, the application provides an office park electric load prediction and dispatching method for a new power system, which has the advantages of solving the problem of fragmentation of the whole chain, realizing panoramic data view, collaborative prediction model, quantitative risk assessment and multi-objective optimization dispatching, thereby improving prediction accuracy and dispatching flexibility, and supporting active participation of office parks in energy management.

[0005] In order to solve the above technical problems, the technical scheme adopted by the application is: The office park electric load prediction and dispatching method for a new type of power system comprises the following steps: Step S1, multi-source data fusion and panoramic view construction, Deploy an intelligent sensor network to collect historical load data in the park in real time; synchronously acquire park operation data, meteorological data and special event calendar; wherein the operation data at least includes the number of settled enterprises, office area utilization rate, equipment account and energy consumption label; align and associate the above multi-source heterogeneous data according to a unified timestamp and spatial label, and construct a time-space consistent park energy data panoramic view; Step S2, data preprocessing and feature directional extraction, The data in the panoramic view is subjected to quality management, including identifying and repairing data anomalies by using an anomaly detection algorithm, and normalizing the multi-source data to eliminate the dimension effect; based on the target of the prediction task, features are extracted from the managed data, including at least time series dynamic features for capturing instantaneous changes and complex dependencies, and structured static features for reflecting operation status and environmental influence; Step S3, multi-time scale prediction based on model cooperation, Construct and cooperate short-term, medium-term and long-term prediction models to form a hierarchical prediction system: a short-term prediction module adopts a time series deep learning model, and the time series dynamic features extracted in the previous step are mainly input, and a high time resolution load prediction curve is output; a medium-term prediction module adopts a decomposition-based regression model, and the structured static features and special event information extracted in the previous step are mainly input, and the seasonality, trend and calendar effect of the load are analyzed and predicted; a long-term prediction module adopts a statistical model combined with external planning scenarios to predict the long-term development trend of the load; wherein the output of the medium-term prediction module provides periodic reference and event correction reference for the short-term prediction module; the output of the long-term prediction module provides macro trend constraint for the medium-term prediction module; Step S4, prediction result optimization and evaluation, The prediction result of step S3 is subjected to uncertainty evaluation, specifically: a probability prediction method is used to output prediction intervals with different confidence intervals, and a random simulation technique is used to quantify the risk probability under extreme scenarios; an online monitoring mechanism of model performance is established, based on rolling time window cross validation and multi-dimensional evaluation indexes, the accuracy of each prediction module is continuously evaluated; when the evaluation result triggers the preset retraining condition, the corresponding module parameter update or structure optimization is automatically started; Step S5, hierarchical dispatching decision and execution based on multi-objective optimization, The dynamic scheduling optimization model is established with the park operation economy, the renewable energy local consumption rate and the power supply reliability as the optimization targets; the deterministic prediction result of step S3 and the uncertainty quantification result of step S4 are jointly used as the input of the optimization model, and the optimal scheduling strategy is solved, and the hierarchical scheduling instructions are generated and executed according to the optimal scheduling strategy, and the instructions include normal optimization scheduling instructions, emergency response scheduling instructions and long-term planning suggestion instructions.

[0006] Further, in step S5, the normal optimization scheduling instructions make the daily rolling output plan of the distributed resources such as photovoltaic and energy storage based on the short-term prediction result; the emergency response scheduling instructions automatically start the demand side response resources and standby energy storage when the deviation or risk probability of the real-time load from the predicted value exceeds the threshold; and the long-term planning suggestion instructions generate the planning scheme of power distribution network expansion and distributed energy configuration based on the medium and long-term prediction result.

[0007] Further, in step S1, the special event calendar at least includes predictable park activities and external events; based on the special event calendar and historical load data, an event-load correlation model is constructed by using knowledge graph technology, the model correlates event attributes with load fluctuation patterns; when predicting, the influence of future planned events is quantified as a load adjustment factor by querying the model, and the load adjustment factor is injected into the prediction module of step S3.

[0008] Further, in step S2, the data quality management further includes a data credibility fusion step: by using the physical or statistical correlation relationship between multiple data sources, cross-validation and credibility weighting are performed by using a belief propagation algorithm, when a certain data source is abnormal, it is repaired or replaced according to other high-credibility data sources.

[0009] Further, in step S3, a cooperative interface is arranged between the short-term prediction module and the medium-term prediction module; the significant calendar effect component and the trend component extracted by the medium-term prediction module are fed into the short-term prediction module in the form of attention weight or feature vector through the cooperative interface, to guide the prediction calibration of the key time point.

[0010] Further, the construction of the short-term prediction module adopts a transfer learning paradigm, specifically including: based on the public or aggregated data of multiple similar parks, a general load prediction base model is pre-trained in the cloud; when deployed in the target park, the base model is fine-tuned with a small amount of local data to quickly adapt to the individual characteristics of the park.

[0011] Further, in step S5, the solving strategy of the dynamic scheduling optimization model is obtained through reinforcement learning training: a virtual simulation environment of the park containing photovoltaic, energy storage and flexible load is constructed; an agent proxy is trained in the environment to predict the result as a state perception, to schedule a dynamic action as a decision output, to take the lowest comprehensive cost and the highest renewable energy consumption as the reward target, and to learn the optimal scheduling strategy through interactive trial and error.

[0012] Further, the method further comprises a carbon flow management and optimization step: based on the real-time scheduling strategy of step S5 and the carbon intensity factor of the power grid, the park power carbon flow atlas is dynamically calculated and generated to realize the carbon emission tracing of the power consumption in each period; the carbon flow atlas is further used as one of the optimization targets to participate in the multi-objective optimization decision of step S5 to realize the low-carbon operation on the basis of economic scheduling.

[0013] Compared with the prior art, the beneficial effects of the present application are: The office park electrical load prediction and scheduling method provided in the present application solves the problem of fragmentation of all links in the background technology by constructing a panoramic view of multi-source data fusion, data preprocessing and feature extraction, multi-time scale prediction, prediction result optimization and evaluation, and hierarchical scheduling decision based on multi-objective optimization, and has the advantages of realizing panoramic data view, collaborative prediction model, quantitative risk assessment and multi-objective optimization scheduling, thereby improving prediction accuracy and scheduling flexibility, and supporting the active participation of office parks in energy management. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is a flowchart of the present application. DETAILED DESCRIPTION

[0015] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined purposes, the specific embodiments, structures, features and effects according to the present application are described in detail below in combination with the preferred embodiments and the drawings.

[0016] It should be noted that in the present application, the orientation words such as "up, down, top, bottom" are generally directed to the directions shown in the drawings, or are directed to the vertical, perpendicular or gravity directions of the components themselves; similarly, for the convenience of understanding and description, "inner, outer" refers to the inner and outer of the contour of each component itself, but the above orientation words are not used to limit the present application.

[0017] EMBODIMENT The application proposes an office park electrical load prediction and dispatching method for new power systems. Through multi-source data fusion and panoramic view construction, efficient use of data is achieved. Through data preprocessing and feature directional extraction, high-quality input is provided for prediction. Through multi-time scale prediction based on model collaboration, the accuracy and robustness of the prediction are improved. Through prediction result optimization and evaluation, the continuous performance of the model is ensured. Finally, through multi-objective optimization-based hierarchical scheduling decision and execution, the economic, efficient and reliable operation of the park energy is realized.

[0018] For ease of understanding, some key terms in the present embodiment are explained as follows: Intelligent sensor network: refers to various sensors and their communication networks widely deployed in the park, which are used to collect real-time and continuous data of power load, environmental parameters, equipment status, etc., providing basic data support for subsequent data analysis and prediction. Park energy data panoramic view: refers to the integration, alignment and association of heterogeneous data from different data sources through unified timestamps and spatial labels, forming a park energy information set that reflects the overall operation status of the park energy system. Data quality governance: refers to a series of processing operations on raw data, including identifying and repairing abnormal values, missing values, noise, etc., and normalizing data of different dimensions to improve data accuracy, integrity and usability, ensuring the quality of data analysis and model training.

[0019] Feature set: refers to a set of input variables that have a significant impact on load prediction, which are extracted from the governed data according to the needs of the prediction task. Time series dynamic feature: refers to data features that can capture the instantaneous characteristics and complex dependence of load changes over time, such as historical load values, load change rates, periodic fluctuations, etc., which are usually extracted through time series analysis methods.

[0020] Structured static feature: refers to data features that reflect the relatively stable or slowly changing factors of park operation status, environmental conditions, etc., which have a long-term or trend impact on load level, such as the number of tenants, office area, temperature, holiday information, etc. Multi-time scale prediction: refers to the construction and collaborative operation of multiple prediction models for different time span load prediction needs, to capture the load variation law at different time scales, forming a hierarchical and complementary prediction system. Dynamic dispatching optimization model: refers to a mathematical optimization model that takes multiple objectives such as park operation economy, local renewable energy consumption rate and power supply reliability as the guide, used to calculate the optimal energy dispatching strategy in real time or quasi-real time considering various constraints.

[0021] The office park electrical load prediction and dispatching method for new power systems of the present application specifically includes the following steps.

[0022] Step S1, multi-source data fusion and panoramic view construction. Deploy intelligent sensor network to collect real-time historical load data of the park. Synchronize to obtain park operation data, meteorological data and special event calendar. Among them, the operation data at least includes the number of settled enterprises, office area utilization rate, equipment account and energy consumption label. The park can manually record the meter data as historical load data periodically, and the number of settled enterprises, office area utilization rate and other operation data are manually counted. Meteorological data can be manually downloaded from a public meteorological website, and special event calendar is recorded through a manually maintained table. These data are then manually input into a data storage system. Align and associate the above multi-source heterogeneous data according to a unified timestamp and spatial label to construct a spatio-temporally consistent park energy data panoramic view. In this application, a relational database can be used to store data from different sources, and database query language can be used for data alignment and association. For example, by matching date and time fields, load data and meteorological data are aligned, and by default, the regional code of the park is used to associate the load data of different regions, thereby forming a preliminary integrated view.

[0023] Step S2, data preprocessing and feature directional extraction. Specifically, the data in the panoramic view are subjected to quality governance, including identifying and repairing data anomalies using anomaly detection algorithms, and normalizing multi-source data to eliminate dimensional effects. As an embodiment, data quality governance can use a statistical threshold-based method to identify abnormal data, for example, data exceeding three times the standard deviation of the historical average is marked as abnormal. For the identified abnormal data, it can be simply replaced by the previous valid data point or the historical average value at that time point. Normalization can use the maximum-minimum normalization method to linearly scale all data to the [0, 1] interval. Further, based on the target of the prediction task, features are extracted from the governed data, including at least: time series dynamic features for capturing instantaneous changes and complex dependencies, and structured static features for reflecting operational status and environmental influences. The time series dynamic features can simply select the historical load values of the previous few hours or days as input. The structured static features can include the current date, day of the week, month, and current temperature, etc. These features are directly input into the prediction model.

[0024] Step S3, multi-time scale prediction based on model collaboration. Construct and collaborate short-term, medium-term and long-term prediction models to form a hierarchical prediction system. As an embodiment, short-term, medium-term and long-term prediction models can be independently constructed, each model is run separately and generates a prediction result. The collaboration of these models can be done through simple rules, for example, the long-term prediction result is the upper limit of the medium-term prediction, and the medium-term prediction result is the reference baseline of the short-term prediction, but there is no deep information interaction between the models.

[0025] The short-term prediction module adopts a time series deep learning model, and the time series dynamic features extracted in the previous step are the main input, and a high time resolution load prediction curve is output. The short-term prediction module can adopt a basic multi-layer perception model, and the time series dynamic features extracted in the previous step are input, and the load prediction value in the future several hours or days is output through the calculation of the multi-layer neural network. The medium-term prediction module adopts a decomposition-based regression model, and the structured static features and special event information extracted in the previous step are the main input, and the seasonality, trend and calendar effect of the load are analyzed and predicted. As an implementation manner, the medium-term prediction module can adopt a simple linear regression model, and the date, month, day of the week and the like are taken as features, and the binary variables of special events such as holidays are manually added to predict the seasonality, trend and calendar effect components of the load. The long-term prediction module adopts a statistical model combined with external planning scenarios to predict the long-term development trend of the load. The long-term prediction module can adopt an exponential smoothing model, and the external planning scenarios such as artificial input park expansion plan and equipment update cycle are combined to predict the load development trend in the future several years or even longer. The output of the medium-term prediction module provides a periodic reference and event correction reference for the short-term prediction module; and the output of the long-term prediction module provides a macro trend constraint for the medium-term prediction module. For example, the daily load curve predicted by the medium-term prediction is taken as the baseline of the short-term prediction, and the short-term prediction is fine-tuned on the baseline. The output of the long-term prediction module can be taken as a fixed constraint condition of the medium-term prediction model, so that the medium-term prediction result will not deviate from the long-term trend.

[0026] Step S4, prediction result optimization and evaluation. The prediction result of step S3 is evaluated for uncertainty, which includes the following contents. (1) A probability prediction method is used to output prediction intervals with different confidence intervals, and a random simulation technique is used to quantify the risk probability under extreme scenarios. The uncertainty evaluation can adopt a historical error statistical method, according to the distribution of historical prediction errors, to calculate the confidence interval of the prediction value. For the risk probability under extreme scenarios, a simple Monte Carlo simulation can be used, assuming that the error follows a certain fixed distribution, and multiple samplings are performed to estimate the risk. (2) An online monitoring mechanism of model performance is established, based on rolling time window cross-validation and multi-dimensional evaluation indexes, to continuously evaluate the accuracy of each prediction module; when the evaluation result triggers the preset retraining condition, the corresponding module parameter update or structure optimization is automatically started. The online monitoring of model performance can set a fixed evaluation period, for example, once a week. The evaluation index can only use the mean absolute error. When the MAE exceeds a fixed threshold for consecutive weeks, the model retraining is triggered, that is, the latest historical data is used to train the model.

[0027] Step S5: Hierarchical Scheduling Decision and Execution Based on Multi-Objective Optimization. A dynamic scheduling optimization model is established with the optimization objectives of park operation economy, local renewable energy absorption rate, and power supply reliability. As one implementation method, a linear programming-based scheduling optimization model can be established, using park operation economy, local renewable energy absorption rate, and power supply reliability as weighted terms in the objective function. These weights can be manually set according to current priorities. The deterministic prediction results from Step S3 and the uncertainty quantification results from Step S4 are used together as inputs to the optimization model to obtain the optimal scheduling strategy. Based on this, hierarchical scheduling instructions are generated and executed, including normal optimization scheduling instructions, emergency response scheduling instructions, and long-term planning recommendation instructions. As one implementation method, the optimization model can primarily rely on the deterministic prediction results from Step S3 for solution. The uncertainty quantification results from Step S4 can serve as auxiliary information; for example, when uncertainty is high, dispatchers can manually adjust the scheduling strategy instead of directly using it as a hard constraint or objective of the optimization model. The scheduling instructions can be generated by the system, manually reviewed, and then issued for execution.

[0028] This method systematically integrates multi-source data to construct a comprehensive energy view and performs refined data governance and feature extraction, providing a high-quality foundation for forecasting. A multi-timescale forecasting system based on model collaboration effectively improves the accuracy and robustness of load forecasting. Combining uncertainty assessment and online monitoring mechanisms ensures continuous optimization of the forecasting model. Ultimately, through multi-objective optimization, hierarchical scheduling decisions are achieved, transforming office parks from passive load points into proactive, flexible, and interactive smart energy nodes with the power grid, significantly improving the economic efficiency of park energy management, renewable energy absorption rate, and power supply reliability.

[0029] This application further proposes, in step S5, a normalized optimization dispatch instruction to formulate a daily rolling output plan for distributed resources such as photovoltaics and energy storage based on short-term forecast results; an emergency response dispatch instruction to automatically activate demand-side response resources and backup energy storage when the deviation between real-time load and forecast value or the risk probability exceeds a threshold; and a long-term planning recommendation instruction to generate a planning scheme for distribution network expansion and distributed energy configuration based on medium- and long-term forecast results.

[0030] Specifically, routine optimization dispatch instructions refer to regular dispatch instructions generated during the park's daily operation to achieve optimal economic efficiency, renewable energy absorption, and power supply reliability. The formulation of these instructions primarily relies on the high-time-resolution load forecast curve output by the short-term forecasting module in the preceding steps. Based on this short-term forecast result, combined with the predicted output of photovoltaic power generation, the current state of charge and charging / discharging capacity of the energy storage system, and the operating characteristics of other distributed power sources, a specific output plan for distributed resources over a future period is dynamically optimized and generated. This plan is typically updated on a rolling basis, such as every 15 minutes, hourly, or every few hours, to adapt to real-time changes in load and renewable energy output.

[0031] Emergency response dispatch instructions are used to address sudden situations or abnormal events that occur during park operations, ensuring the continuity and stability of power supply. When online monitoring detects a deviation between real-time load and short-term forecasts, or when uncertainty assessments indicate a risk probability exceeding a preset safety threshold, the system will automatically trigger this instruction. This instruction will rapidly activate demand-side response resources within the park to reduce or shift load, and instruct the backup energy storage system to rapidly charge and discharge, quickly filling supply-demand gaps or smoothing system fluctuations.

[0032] The generation of long-term planning recommendations is primarily based on load development trends, seasonal characteristics, calendar effects, and macroeconomic constraints output by the medium- and long-term forecasting module. Based on these medium- and long-term forecasts, combined with the park's future development plan, policy guidance, and techno-economic analysis, the system proposes suggestions for expanding and upgrading the distribution network infrastructure and configuring distributed energy resources. This guidance provides decision support for the strategic investment and construction of the park's energy system, ensuring that the park's energy system can meet future load growth demands and continuously improve the utilization rate of renewable energy and power supply reliability.

[0033] In step S1, the special event calendar includes at least foreseeable park activities and external events; based on the special event calendar and historical load data, an event-load correlation model is constructed using knowledge graph technology, which associates event attributes with load fluctuation patterns; when making predictions, the impact of future planned events is quantified into load adjustment factors by querying the model and injected into the prediction module in step S3.

[0034] External events refer to events occurring outside the park that affect the park's load, such as power outage notices in the surrounding area, large-scale public transportation adjustments, and regional holiday arrangements. To effectively utilize this special event information, this application employs knowledge graph technology to construct an event-load association model. Various attributes of events, such as event type, scale, duration, number of participants, or affected area, are treated as entity or attribute nodes. Load fluctuation patterns, including load peak increments, trough decreases, fluctuation duration, and load curve shape changes, are treated as another type of entity or attribute node. These are connected through semantic relationships such as "cause," "affect," and "have attributes." For example, a "large conference" event entity can be associated with the "number of participants" attribute, and then, through the "cause" relationship, with the load fluctuation pattern entity "load peak increase of 15%." The construction process can utilize historical event records, historical load data, and domain expert knowledge, achieved through technologies such as entity recognition, relation extraction, and knowledge fusion. Through the aforementioned knowledge graph, the detailed attributes of specific events can be accurately mapped and associated with specific load fluctuation patterns observed in historical load data. For example, when an event is identified as “a department holds a half-day internal training”, the model can query the corresponding historical load fluctuation pattern based on its attributes such as “training” type, “half-day” duration, and “internal” impact range, such as “the load drops slightly at noon and returns to normal in the afternoon”.

[0035] When forecasting, if a planned special event, such as an upcoming holiday or scheduled equipment maintenance, is identified, the system queries the built event-load correlation model. Based on the specific attributes of the event, the model outputs a quantified load adjustment factor. This factor can be a time-series load increment or decrement curve, a multiplicative correction factor, or a function describing changes in the shape of the load curve. For example, for an event expected to increase pedestrian traffic in the park, the model might output an adjustment factor that increases the load by 100kW during a specific period of the event. This quantified adjustment factor is then injected into the forecasting module in step S3. It can serve as an additional input feature to the short-term forecasting module or be used to directly correct the baseline load curve generated by the medium-term forecasting module, thereby enabling the forecast results to more accurately reflect the expected impact of future special events on the load.

[0036] In step S2, the data quality governance also includes a data credibility fusion step: using the physical or statistical correlation between multi-source data, cross-validation and credibility weighting are performed through the belief propagation algorithm. When a data source is abnormal, it is repaired or replaced based on other high-credibility data sources.

[0037] The core of data credibility fusion lies in the ability to correct or replace a data source with other related and highly credible data sources when a data source is abnormal or its data quality is questionable, thereby ensuring that the data input into the prediction model has higher accuracy and robustness.

[0038] Physical correlations refer to the inherent connections between different data sources based on physical laws or system structures. For example, load data from different areas within a park may exhibit spatial correlations; a direct physical causal relationship exists between photovoltaic power generation and solar radiation intensity data. Statistical correlations refer to indirect but significant correlations between different data sources discovered through statistical analysis. For example, there is a strong statistical correlation between air conditioning load in office buildings and outdoor temperature and humidity.

[0039] In this step, the confidence propagation algorithm is used to propagate the confidence information of each data source in the constructed multi-source data association network. Once the initial confidence of a data source is evaluated, this confidence information is propagated along the edges of the graph to its associated nodes and updated according to the association strength. Through multiple iterative propagations, the final confidence of each data source converges to a stable value, reflecting its overall credibility within the entire data network.

[0040] When data from a temperature sensor shows an anomaly, it can be compared with data from other nearby temperature sensors or with temperature values ​​calculated based on a physical model to determine the authenticity of the anomaly. This is the cross-validation referred to in this application. Credibility weighting means that data sources with higher credibility have a greater influence during data fusion or repair, while data sources with lower credibility have less influence. This weighting mechanism ensures that the system prioritizes more reliable information during the data fusion process.

[0041] When a data source is determined to be abnormal or its credibility falls below a preset threshold, the system will initiate a repair or replacement mechanism. This involves estimating and replacing outlier values ​​through interpolation, regression models, or calculations based on physical models. Additionally, when a sensor completely fails, it can be switched to its redundant sensor or replaced using model predictions.

[0042] In step S3, a collaborative interface is provided between the short-term prediction module and the medium-term prediction module; the significant calendar effect component and trend component extracted by the medium-term prediction module are fed into the short-term prediction module in the form of attention weights or feature vectors through the collaborative interface to guide it in making prediction calibrations for key time points.

[0043] Specifically, when performing load forecasting, the medium-term forecasting module typically employs a decomposition-based regression model to break down the original load sequence into multiple interpretable components. The significant calendar effect component refers to load fluctuation patterns caused by periodic factors such as dates, days of the week, and holidays, such as load differences between weekdays and weekends, or load peaks or troughs during specific holidays. The trend component reflects the long-term directionality of load changes over time, such as slow increases or decreases in load due to park development, equipment upgrades, or energy-saving measures. These components are usually extracted in the form of time series data or their statistical characteristics.

[0044] To enable the short-term forecasting module to effectively utilize this intermediate-term information, the collaborative interface supports data feed in two main forms. The first is attention weights, where the intermediate-term forecasting module transforms calendar effect components and trend components into a set of weights. These weights indicate to the short-term forecasting module which time points or features to allocate more attention to when processing its own input data. For example, when a holiday is predicted to cause a significant drop in load, the intermediate-term module can generate corresponding attention weights, allowing the short-term model to more strongly focus on and adjust its output to reflect this downward trend when forecasting load during that period. The second form is feature vectors, where the intermediate-term forecasting module encodes calendar effect components and trend components into numerical feature vectors. These feature vectors can serve as additional input features for the short-term forecasting module, fed into the deep learning model along with the original time-series dynamic features. The short-term forecasting module learns the complex relationship between these feature vectors and load, thereby internally adjusting its forecasting logic.

[0045] The short-term prediction module is constructed using a transfer learning paradigm, specifically including: pre-training a general load prediction basic model in the cloud based on public or aggregated data from multiple similar parks; and fine-tuning the basic model using a small amount of local data when deploying it in the target park to quickly adapt to the park's personalized characteristics.

[0046] Specifically, based on publicly available or aggregated data from multiple similar office parks, a general load forecasting model is pre-trained in the cloud to form an initial model with good generalization capabilities. In this embodiment, historical load data, meteorological data, and operational data from multiple office parks with similar functions, sizes, or geographical locations can be collected, cleaned, and integrated. On a high-performance computing platform, a general deep learning model, including Long Short-Term Memory networks, gated recurrent units, or Transformer time-series deep learning models, is trained using these multi-source data to identify common patterns of load changes in different parks, such as weekdays versus weekends, seasonal variations, and holiday effects.

[0047] When deploying in a target park, the base model is fine-tuned using a small amount of local data to quickly adapt to the park's unique characteristics. Since each park differs in building structure, equipment configuration, and personnel activity patterns, its load patterns are also unique. Therefore, only a small amount of historical load data and relevant feature data from the target park itself needs to be collected to retrain and fine-tune specific or all layers of the pre-trained base model with a small learning rate. This fine-tuning process is typically much faster than training the model from scratch and requires less local historical data, enabling the model to accurately capture the unique load patterns and operational habits of the target park, thereby improving prediction accuracy.

[0048] In step S5, the solution strategy of the dynamic scheduling optimization model is obtained through reinforcement learning training: a virtual simulation environment of the park including photovoltaic, energy storage and flexible load is constructed; the intelligent agent is trained in the environment, taking the prediction result as the state perception, the scheduling action as the decision output, and the lowest comprehensive cost and the highest renewable energy consumption as the reward objectives, and learns the optimal scheduling strategy through interactive trial and error.

[0049] The method also includes carbon flow management and optimization steps: based on the real-time scheduling strategy and power grid carbon intensity factor in step S5, the park's power carbon flow map is dynamically calculated and generated to trace the source of carbon emissions from electricity consumption in each time period; the carbon flow map is further used as one of the optimization objectives to participate in the multi-objective optimization decision in step S5 to achieve low-carbon operation based on economic dispatch.

[0050] The real-time dispatch strategy refers to the optimal dispatch strategy obtained by solving the dynamic dispatch optimization model in step S5 above, which includes normal optimization dispatch instructions, emergency response dispatch instructions, and long-term planning suggestion instructions. These strategies determine the operating status and power flow of various energy devices such as photovoltaic, energy storage, and flexible loads within the park. The grid carbon intensity factor refers to the carbon emissions generated per unit of electricity. This factor can change in real time, reflecting the instantaneous combination of thermal power, hydropower, wind power, and photovoltaic power sources in the grid. For example, when the proportion of renewable energy generation in the grid is high, the carbon intensity factor is low; conversely, it is high. Dynamically calculating and generating the park's electricity carbon flow map means calculating the park's carbon emissions at different time periods by combining the electricity consumption determined by the real-time dispatch strategy and the grid carbon intensity factor. If the park purchases 100 kWh of electricity from the grid at a certain time period, and the grid carbon intensity factor at that time is 0.5 kgCO... 2 If the carbon emission during that period is 50 kg CO, then the carbon emission during that period is / kWh. 2A park's electricity carbon flow map is a visual or data-driven representation that shows the carbon emissions corresponding to electricity consumption at different loads and time periods within the park, as well as the sources of these emissions. It can be in the form of a time series graph, heat map, or network diagram, intuitively reflecting the park's carbon emission status. Through the electricity carbon flow map, it is possible to clearly trace how much carbon emissions were generated by the park's electricity consumption within a specific time period, and the specific sources of these emissions.

[0051] The carbon flow map is further incorporated as one of the optimization objectives in the multi-objective optimization decision-making process of step S5. When solving the dynamic scheduling optimization model, in addition to considering the original optimization objectives such as park operation economy, local renewable energy absorption rate, and power supply reliability, the model also considers carbon emission objectives, such as minimizing total carbon emissions. The optimization algorithm seeks a scheduling strategy that minimizes carbon emissions while satisfying other constraints. This may involve strategies such as reducing power purchases from the grid when grid carbon intensity is high, increasing local renewable energy absorption, or activating energy storage discharge. The goal is to achieve low-carbon operation based on economic scheduling, emphasizing that while pursuing low-carbon development, the original economic objectives should not be abandoned or severely sacrificed.

[0052] The above technical solution introduces carbon flow management and optimization steps into the existing multi-objective optimization scheduling strategy, which considers factors such as the economic efficiency of park operation, local renewable energy absorption rate, and power supply reliability. By dynamically calculating and generating a carbon flow map of the park's electricity, the carbon emission sources of electricity consumption at different times can be clearly traced, allowing park managers to intuitively understand their carbon footprint. More importantly, by incorporating the carbon flow map as a new optimization objective into the dynamic scheduling optimization decision-making process, the scheduling strategy not only considers economic benefits and power supply reliability but also proactively guides the park to use electricity during periods of lower grid carbon intensity. This effectively promotes the low-carbon operation of the park based on economic scheduling, achieving a synergistic improvement in both environmental and economic benefits.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for predicting and dispatching electrical load in office parks for new power systems, characterized in that: Including the following steps: Step S1: Multi-source data fusion and panoramic view construction. Deploy an intelligent sensor network to collect historical load data of the park in real time; synchronously acquire park operation data, meteorological data and special event calendars; wherein, the operation data includes at least the number of resident enterprises, office space utilization rate, equipment ledger and energy consumption labels; align and associate the above multi-source heterogeneous data with unified timestamps and spatial labels to construct a spatiotemporally consistent panoramic view of park energy data; Step S2: Data preprocessing and targeted feature extraction. The data in the panoramic view is subjected to quality governance, including the use of anomaly detection algorithms to identify and repair data anomalies, and the normalization of multi-source data to eliminate the influence of units; based on the goal of the prediction task, a feature set is extracted from the governed data in a targeted manner, the feature set including at least: time-series dynamic features for capturing instantaneous changes and complex dependencies, and structured static features for reflecting operational status and environmental impact; Step S3: Multi-timescale prediction based on model collaboration. A hierarchical forecasting system is formed by constructing and coordinating short-term, medium-term, and long-term forecasting models: The short-term forecasting module uses a time-series deep learning model, taking the time-series dynamic features extracted in the preceding steps as the main input, and outputting a high-time-resolution load forecasting curve; the medium-term forecasting module uses a decomposition-based regression model, taking the structured static features extracted in the preceding steps and special event information as the main input, and analyzing and predicting the seasonality, trend, and calendar effects of the load; the long-term forecasting module uses a statistical model combined with external planning scenarios to predict the long-term development trend of the load. The output of the medium-term forecasting module provides a periodic benchmark and event correction reference for the short-term forecasting module; the output of the long-term forecasting module provides macroeconomic trend constraints for the medium-term forecasting module. Step S4: Optimization and evaluation of prediction results. Uncertainty assessment is performed on the prediction results of step S3, specifically: the prediction intervals with different confidence intervals are output using a probabilistic prediction method, and the risk probability under extreme scenarios is quantified using stochastic simulation technology; an online monitoring mechanism for model performance is established, and the accuracy of each prediction module is continuously evaluated based on rolling time window cross-validation and multi-dimensional evaluation indicators; when the evaluation results trigger the preset retraining conditions, the parameter update or structure optimization of the corresponding module is automatically started. Step S5: Hierarchical scheduling decision and execution based on multi-objective optimization. With the optimization objectives of park operation economy, local renewable energy absorption rate and power supply reliability, a dynamic scheduling optimization model is established. The deterministic prediction results of step S3 and the uncertainty quantification results of step S4 are used as inputs to the optimization model to obtain the optimal scheduling strategy. Based on this, hierarchical scheduling instructions are generated and executed, including normal optimization scheduling instructions, emergency response scheduling instructions and long-term planning suggestion instructions.

2. The method for predicting and dispatching power load in office parks for new power systems according to claim 1, characterized in that, In step S5, the routine optimization dispatch instruction formulates a daily rolling output plan for distributed resources such as photovoltaics and energy storage based on short-term forecast results; the emergency response dispatch instruction automatically activates demand-side response resources and backup energy storage when the deviation between real-time load and forecast value or the risk probability exceeds a threshold; and the long-term planning suggestion instruction generates a planning scheme for distribution network expansion and distributed energy configuration based on medium- and long-term forecast results.

3. The method for predicting and dispatching power load in office parks for new power systems according to claim 2, characterized in that, In step S1, the special event calendar includes at least foreseeable park activities and external events; Based on the special event calendar and historical load data, an event-load correlation model is constructed using knowledge graph technology. The model associates event attributes with load fluctuation patterns. When making predictions, the impact of future planned events is quantified into load adjustment factors by querying the model and injected into the prediction module in step S3.

4. The method for predicting and dispatching power load in office parks for new power systems according to claim 3, characterized in that, In step S2, the data quality governance also includes a data credibility fusion step: using the physical or statistical correlation between multi-source data, cross-validation and credibility weighting are performed through the belief propagation algorithm. When a data source is abnormal, it is repaired or replaced based on other high-credibility data sources.

5. The method for predicting and dispatching power load in office parks for new power systems according to claim 4, characterized in that, In step S3, a collaborative interface is provided between the short-term prediction module and the medium-term prediction module; the significant calendar effect component and trend component extracted by the medium-term prediction module are fed into the short-term prediction module in the form of attention weights or feature vectors through the collaborative interface to guide it in making prediction calibrations for key time points.

6. The method for predicting and dispatching power load in office parks for new power systems according to claim 5, characterized in that, The short-term prediction module is constructed using a transfer learning paradigm, specifically including: pre-training a general load prediction basic model in the cloud based on public or aggregated data from multiple similar parks; and fine-tuning the basic model using a small amount of local data when deploying it in the target park to quickly adapt to the park's personalized characteristics.

7. The method for predicting and dispatching power load in office parks for new power systems according to claim 6, characterized in that, In step S5, the solution strategy of the dynamic scheduling optimization model is obtained through reinforcement learning training: a virtual simulation environment of the park including photovoltaic, energy storage and flexible load is constructed; the intelligent agent is trained in the environment, taking the prediction result as the state perception, the scheduling action as the decision output, and the lowest comprehensive cost and the highest renewable energy consumption as the reward objectives, and learns the optimal scheduling strategy through interactive trial and error.

8. The method for predicting and dispatching power load in office parks for new power systems according to claim 7, characterized in that, The method also includes a carbon flow management and optimization step: based on the real-time scheduling strategy and power grid carbon intensity factor in step S5, the park's power carbon flow map is dynamically calculated and generated to achieve carbon emission traceability for electricity consumption in each time period. The carbon flow map is further used as one of the optimization objectives in the multi-objective optimization decision-making of step S5, so as to achieve low-carbon operation based on economic scheduling.

Images