Load prediction and energy storage scheduling method and system for algorithm and electricity collaborative scene

Abstract

The application relates to a load prediction and energy storage scheduling method and system for an algorithm and electricity collaborative scene. A load prediction model is used on the prediction side to output predicted load, prediction uncertainty and key period risk score; on the scheduling side, the prediction uncertainty and the key period risk score are mapped into discharge power, state of charge and risk penalty constraint, and a two-stage energy storage optimization model is used to generate a charge and discharge plan; thus, the prediction risk is directly involved in the energy storage constraint, and the scheduling result and the income result adopt consistent data caliber.

Description

Technical Field

[0001] This application relates to the fields of load forecasting, energy storage control and data processing technology, and in particular to a load forecasting and energy storage scheduling method and system for computing and power coordination scenarios. Background Technology

[0002] In user-side energy scenarios such as industrial and commercial parks, buildings, energy storage stations, microgrids, and virtual power plants, load forecasting, energy storage dispatching, and revenue calculation are typically handled by different systems or algorithms. The forecasting system outputs future load curves, the energy storage dispatching system generates charging and discharging strategies based on electricity prices or human experience, and the revenue system separately calculates actual charging and discharging revenue. Due to inconsistencies in data definitions, time granularity, node identifiers, and financial calculation rules among these systems, business personnel often need to manually convert data between multiple systems to determine the predicted load, dispatching plan, and peak-valley arbitrage revenue for a specific node within a given date or period.

[0003] The existing solutions have at least the following problems: 1. Disconnect between forecasting and dispatching. Load forecasting results cannot be directly converted into energy storage charging and discharging actions, and dispatching strategies also struggle to utilize the latest forecast curves and electricity price information in a timely manner.

[0004] 2. Difficulty in aligning multi-source data. Actual load, forecast load, information technology (IT) load, electricity price, special electricity price for holidays, and planned power may come from different data tables or time-series systems, with different time granularities such as 5 minutes, 15 minutes, and 1 hour. IT load refers to the total active power consumed by information technology equipment (servers, storage, network equipment, etc.) in the data center. It is a core component of data center energy consumption and directly determines the capacity design of the cooling system and power supply and distribution system.

[0005] 3. Instability in revenue calculation. If large language model agents or manual reports are decoupled from the business system and the returned fields are recalculated to determine electricity prices, power, and electricity consumption, inconsistencies in financial fields or unverifiable results can easily occur.

[0006] 4. Lack of domain constraints in agent tool calls. When only general interfaces are provided, agents may incorrectly select load forecasting, IT load forecasting, energy storage scheduling, or revenue query tools, mix new and old forecasting models, or generate unreliable results when nodes, dates, or granularities are missing. Summary of the Invention

[0007] This application proposes a load forecasting and energy storage scheduling method and system for computing and power collaboration scenarios, which is used to solve at least one of the above-mentioned technical problems.

[0008] To achieve the above objectives, this application adopts the following technical solution: Firstly, a load forecasting and energy storage scheduling method for computing-power collaborative scenarios is provided, including: On the forecasting side, load forecasting data is obtained, which includes continuous variables and categorical features. The continuous variables include electricity load, IT load, and electricity price. The categorical features include node type, time period type identifier, scenario label, critical period label, and energy storage device identifier. The trained load forecasting model is used to process the load forecasting data to obtain load forecasting results. The load forecasting results include forecasted load, forecast uncertainty measure, and critical period risk score. The forecast uncertainty measure is used to describe the forecast fluctuation or confidence interval width at future times, and the critical period risk score is used to characterize the scheduling risk during critical periods. The load prediction model comprises an input embedding layer, a short-scale branch, a long-scale branch, a scene gating layer, a key period enhancement layer, and an output layer. The input embedding layer concatenates the continuous variables and categorical features to obtain concatenated features. The short-scale branch captures rapid fluctuations in the concatenated features over short time periods to obtain short-scale features. The long-scale branch captures trends in the concatenated features over long time periods to obtain long-scale features. The scene gating layer performs weighted fusion of the short-scale and long-scale features based on gating weights to obtain fused features, where the gating weights are determined based on the scene labels. The key period enhancement layer enhances the fused features using an attention mechanism to obtain enhanced features, where the attention score is defined by an enhancement coefficient determined by the key period labels. The output layer outputs the load prediction result based on the enhanced features. On the dispatching side, the load forecast results are obtained; using the constructed energy storage optimization model, a charging and discharging plan for the energy storage equipment is generated based on the load forecast results and electricity prices; wherein, in the first stage, the forecast uncertainty metric and the critical period risk score are mapped to discharge power constraints, state of charge constraints and risk penalty weights, and a baseline solution that satisfies the constraints is generated; in the second stage, an improved genetic algorithm is used to search in the neighborhood of the baseline solution to ensure that the candidate solutions continue to satisfy the constraints, and the final solution is obtained.

[0009] In one possible design approach of the first aspect, the critical time period includes: peak or high-peak periods, peak-valley or electricity price switching boundaries, and / or periods with a high incidence of historical errors; the risk score of the critical time period is defined by the prediction uncertainty measure, peak or high-peak period identifier, peak-valley or electricity price switching boundary identifier, and / or period with a high incidence of historical errors identifier.

[0010] In one possible design approach of the first aspect, the short-scale branch employs one-dimensional convolution and short data block encoding, while the long-scale branch employs long data block encoding and Transformer encoding.

[0011] In one possible design approach for the first aspect, To query the time index, Key moment index; basic attention score Enhancement items during key periods Revised to The key period enhancement is determined by a weighted sum of peak or high-peak period identifiers, peak-valley or electricity price switching boundary identifiers, and historical error-prone period identifiers: in, This is an indicator for peak or high-peak hours. This serves as a boundary marker for peak / valley or electricity price switching. This serves as a marker for periods with a high incidence of historical errors. , and For weighted weights.

[0012] In one possible design approach of the first aspect, the loss function of the load forecasting model for: in, This represents the prediction error for all time points. This is the weighted error for the critical period. For uncertainty calibration items, To monitor losses based on risk scores, , , For weights.

[0013] In one possible design approach of the first aspect, the constraints include: a conservative upper limit on the load. Discharge power contraction coefficient SOC conservative margin and risk penalty weight ,in: in, For a moment Point-based load forecasting For a moment The measure of prediction uncertainty For a moment Risk score during key periods , , , , , , , For mapping coefficients, This is the minimum discharge ratio. It is the basic SOC safety margin reserved when both the uncertainty measure and the risk score are 0, and max means to calculate the maximum value.

[0014] In one possible design approach of the first aspect, an improved genetic algorithm is used to optimize based on the baseline solution, specifically including: Set up a chromosome and initialize a population. The chromosome is set as an action segment code consisting of action type, power level and duration. The initial population includes: the benchmark solution and the neighborhood perturbation solution of the benchmark solution. The fitness calculation, selection, crossover, mutation, constraint repair, and elite retention processes are executed sequentially until convergence.

[0015] In one possible design approach of the first aspect, the fitness F Expressed as: In the formula, the basic income item This represents the difference between discharge revenue and charging cost, and includes a risk penalty item. This indicates the penalty for exceeding the allowable discharge power during high-risk periods; action switching penalty item. Used to limit frequent charge / discharge switching and constrain repair cost items. Record the adjustment amount required to restore the feasibility of candidate solutions; , and The weights are non-negative.

[0016] Secondly, a load forecasting and energy storage scheduling system for computing and power collaboration scenarios is provided, including one or more functional units or modules for performing the method described in the first aspect.

[0017] Thirdly, an electronic device is provided, comprising: a processor, and a memory coupled to the processor, the memory for storing a computer program; the processor for executing the computer program stored in the memory such that the electronic device performs the method as described in any possible implementation of the first aspect.

[0018] Beneficial effects: Based on the above technical solution, in addition to outputting the predicted values ​​at the output points, the load forecasting model also outputs uncertainty metrics and risk scores for critical periods. The risk information participates in the energy storage power and state of charge constraints from the first stage, and continues to be used for candidate solution screening and repair in the second stage. This can reduce the data conversion error between forecasting and scheduling, and improve the executability of energy storage plans and the verifiability of the benefit results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the core process of load forecasting and energy storage scheduling provided in Embodiment 1 of this application; Figure 2 This is a schematic diagram of the load forecasting model provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification and the above-mentioned figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0024] Example 1 like Figure 1 As shown, the core process of load forecasting and energy storage scheduling in this embodiment includes: Step S101, on the forecasting side, load forecasting data is obtained. The load forecasting data includes continuous variables and categorical features. The continuous variables include electricity load, information technology (IT) load, and electricity price. The categorical features include node type, time period type identifier, scenario label, critical period label, and energy storage device identifier. The trained load forecasting model is used to process the load forecasting data to obtain load forecasting results. The load forecasting results include forecasted load, forecast uncertainty measure, and critical period risk score. The forecast uncertainty measure is used to describe the forecast fluctuation or confidence interval width at future times. The critical period risk score is used to characterize the scheduling risk during critical periods. Among them, such as Figure 2 As shown, the load prediction model includes: an input embedding layer 201, a short-scale branch 202, a long-scale branch 203, a scene gating layer 204, a key period enhancement layer 205, and an output layer 206. The input embedding layer 201 concatenates the continuous variables and the categorical features to obtain concatenated features. The short-scale branch 202 captures rapid fluctuations in the concatenated features over short time periods to obtain short-scale features. The long-scale branch 203 captures trends in the concatenated features over long time periods to obtain long-scale features. The scene gating layer 204 performs weighted fusion of the short-scale and long-scale features based on gating weights to obtain fused features, where the gating weights are determined based on the scene labels. The key period enhancement layer 205 enhances the fused features using an attention mechanism to obtain enhanced features, where the attention score is defined by an enhancement coefficient determined by the key period labels. The output layer 206 outputs the load prediction result based on the enhanced features. Step S102: The scheduling side obtains the load forecast result; using the constructed energy storage optimization model, a charging and discharging plan for the energy storage device is generated based on the load forecast result and the electricity price; wherein, in the first stage, the forecast uncertainty metric and the critical period risk score are mapped to discharge power constraints, state of charge constraints and risk penalty weights, and a baseline solution that satisfies the constraints is generated by combining the electricity price and energy storage device parameters; in the second stage, an improved genetic algorithm is used to search in the neighborhood of the baseline solution so that the candidate solution continues to satisfy the constraints, and the final solution is obtained.

[0025] In one possible implementation, the critical time period includes: peak or high-peak periods, peak-valley or electricity price switching boundaries, and / or periods with a high incidence of historical errors; the risk score of the critical time period is defined by the prediction uncertainty measure, peak or high-peak period identifier, peak-valley or electricity price switching boundary identifier, and / or period with a high incidence of historical errors identifier.

[0026] In one possible implementation, the short-scale branch 202 employs one-dimensional convolution and short data block (Patch) encoding, while the long-scale branch 203 employs long data block encoding and Transformer encoding.

[0027] In one possible implementation, To query the time index, Key moment index; basic attention score Enhancement items during key periods Revised to The key period enhancement is determined by a weighted sum of peak or high-peak period identifiers, peak-valley or electricity price switching boundary identifiers, and historical error-prone period identifiers: in, This is an indicator for peak or high-peak hours. This serves as a boundary marker for peak / valley or electricity price switching. This serves as a marker for periods with a high incidence of historical errors. , and For weighted weights.

[0028] In one possible implementation, the loss function of the load forecasting model for: in, This represents the prediction error for all time points. This is the weighted error for the critical period. For uncertainty calibration items, To monitor losses based on risk scores, , , For weights.

[0029] In one possible implementation, the above constraints include: a conservative upper bound on the load. Discharge power contraction coefficient SOC conservative margin and risk penalty weight ,in: in, For a moment Point-based load forecasting For a moment The measure of prediction uncertainty For a moment Risk score during key periods , , , , , , , For mapping coefficients, This is the minimum discharge ratio.

[0030] In one possible implementation, an improved genetic algorithm is used to optimize the baseline solution, specifically including: Set up a chromosome and initialize a population. The chromosome is set as an action segment code consisting of action type, power level and duration. The initial population includes: the benchmark solution and the neighborhood perturbation solution of the benchmark solution. The fitness calculation, selection, crossover, mutation, constraint repair, and elite retention processes are executed sequentially until convergence.

[0031] Candidate solutions can also record their offset relative to the baseline solution, which is used to limit the search range.

[0032] In one possible implementation, the fitness is expressed as: In the formula, the basic income item This represents the difference between discharge revenue and charging cost, and includes a risk penalty item. This indicates the penalty for exceeding the allowable discharge power during high-risk periods; action switching penalty item. Used to limit frequent charge / discharge switching and constrain repair cost items. Record the adjustment amount required to restore the feasibility of candidate solutions; , and The weights are non-negative.

[0033] This embodiment incorporates the measurement of prediction uncertainty and the risk score for critical periods into the energy storage constraints, enabling the predicted risk to directly affect the charging and discharging plan; at the same time, it uses a unified time granularity, electricity price, and monetary accuracy to verify the planned revenue and actual revenue, reducing data caliber differences caused by repeated calculations across systems.

[0034] Example 2 This embodiment illustrates how the business system receives load forecast data and energy storage optimization strategies. The system uses the Model Context Protocol (MCP) tool contract to define request parameters and return fields, and performs unified processing on nodes, time windows, data granularity, and financial metrics, outputting forecast curves, scheduling strategies, revenue results, and reports.

[0035] The above processing reduces manual data conversion between forecasting, scheduling, and revenue calculation, and restricts agents from automatically recreating or recalculating business results, thereby improving the traceability of scheduling and revenue results.

[0036] 1. Overall Technical Solution This example mainly provides the following: (1) Receiving business requests. The system receives requests from MCP tools or Hypertext Transfer Protocol (HTTP) debugging interfaces, routes the natural language intent to tool calls in fields such as node load forecasting, IT load forecasting, energy storage scheduling strategies or energy storage revenue queries based on the tool description and parameter contract, and parses parameter slots such as node code, target date, target period, historical data file address and return granularity.

[0037] (2) Unify business parameters. Map the node code to the internal node identifier (ID) and parent node ID, normalize the return granularity to 5 minutes, 15 minutes or 1 hour, and convert the target date to a full-day window or a rolling forecast window.

[0038] (3) Generate forecast data. The algorithm model generates user-side node load forecast curves, IT load forecast curves, or new model forecast curves based on inputs such as historical load, IT load, date type, electricity price information, holiday information, and node characteristics.

[0039] (4) Generate energy storage optimization strategies. The algorithm model or optimization module takes the predicted load, electricity price, energy storage device constraints and revenue target as inputs to generate energy storage charging, discharging or standby strategies and corresponding power for multiple time periods within a day.

[0040] (5) Multi-source data alignment. The business system reads data from the forecast data table, actual time-series data, planned power table, basic time-of-use electricity price table and holiday special electricity price table, and samples or expands the data at a uniform granularity.

[0041] (6) Calculation of indicators and revenue. The system calculates total electricity consumption, peak power, peak time, prediction deviation, prediction accuracy, charging cost, discharge revenue, daily net revenue, actual revenue and planned revenue.

[0042] (7) Output of results. The system returns structured JSON results through the MCP tool and can include a report download address; the prediction curve, strategy table, error index and financial amount are all generated by business services or algorithm results. The agent only generates charts, summaries or scheduling suggestions based on trusted fields and does not perform unconstrained secondary inference on financial fields.

[0043] 2. System Composition 2.1 Intelligent Agent Tool Entry Module This module constitutes the energy domain tool orchestration layer for large-scale model agents. The system encapsulates energy business capabilities into structured tools that can be recognized, selected, and invoked by large models through the MCP tool protocol. Each tool includes a tool name, semantic description, parameter description, required constraints, optional parameters, default granularity, and business security requirements. Large-scale model agents do not directly access the underlying database or generate scheduling results themselves. Instead, they select constrained domain tools based on the user's natural language intent, and then the business service completes parameter normalization, data querying, algorithm result reception, and financial accounting.

[0044] This module should include at least the following tool contracts: (1) getNodeLoadPrediction: User-side node load prediction tool. When a user queries the future electricity consumption curve of a certain industrial or commercial user, park, building, microgrid, energy storage station or virtual power plant node, the agent will route the intent to this tool and pass in the target date, node code, historical load file address and return granularity.

[0045] (2) getNodeLoadPredictionV1: A new model node load prediction tool. This tool maintains the same input parameters, output parameters, and prompt words as the original prediction tool, but receives the results of the new algorithm model through an independent data source. The agent must call this tool when the user explicitly requests it or when the business scenario clearly requires the new model, thereby avoiding the automatic mixing of prediction results from the original model and the new model.

[0046] (3) getITLoadPrediction: Information technology load prediction tool. This tool is designed for IT load capacity analysis and scheduling assistance scenarios. It supports querying for a specified date and also supports scrolling window querying when the target date is empty. It returns the prediction curve, the actual curve, the prediction deviation, and the accuracy.

[0047] (4) getDispatchStrategy: Energy storage dispatch strategy and financial calculation tool. This tool targets semantic intents such as "regulation strategy", "charge and discharge plan", "energy storage arbitrage", and "revenue forecast", and returns the total charging cost, total discharging revenue, daily net revenue, charging and discharging time period, and a table of action instructions for the whole day. The tool contract requires the agent to use the financial amount field returned by the tool when generating the final report, and prohibits unconstrained secondary calculations on electricity price, power and location.

[0048] (5) getStorageArbitrageRevenue: Energy storage peak-valley arbitrage profit tool. This tool is for querying actual profit, planned profit, charging cost, discharge profit and monthly / daily arbitrage profit, and supports outputting auditable profit fields by daily or monthly target period.

[0049] The tool entry layer sequentially completes intent recognition, tool selection, parameter validation, business service invocation, and structured result return. Relative dates are converted to standard date formats; missing nodes are filled in using context, and if uncertain, the caller is asked to supplement them; when granularity is missing, the business side uses the default 5-minute granularity; financial fields are calculated by the business service, and the intelligent agent is only used for interpretation and presentation.

[0050] 2.2 Request Normalization and Node Mapping Module This module converts external request parameters into internal data query parameters. Node codes can be business codes such as snd1, snd2, etc., which the system maps to node ID, parent node ID, line name, and node display name. The target date can be converted into a full-day query window; when the IT load forecast does not specify a target date, the system rounds down to the previous 5 minutes from the current time, forming a scrolling window of the actual load for the previous 12 hours and the predicted load for the next 24 hours.

[0051] Unified support for return granularity: (1) 5-minute granularity, corresponding to 288 points per day.

[0052] (2) 15-minute granularity, corresponding to 96 points per day.

[0053] (3) 1-hour granularity, corresponding to 24 points per day.

[0054] When no granularity is specified in the request, a 5-minute granularity is used by default.

[0055] 2.3 Prediction Data Generation and Fusion Module This module receives the prediction data generated by the algorithm and integrates it with the business side. Ordinary node load prediction data can be obtained from the prediction value field in `ai_load_forecasting`, while new model prediction data can be obtained from prediction results with the same structure via an independent data source. IT load prediction data can be obtained from the prediction curve in the `cscursor` table where `data_name=ITForecastLoad`.

[0056] For predicted curves, the system generates continuous points within a complete time window; for missing predicted points, they are filled in according to business configuration to avoid curve interruptions. For actual curves, the system only retains historical points actually returned by the time series system to avoid mistakenly displaying future uncollected points as 0.

[0057] 2.4 Energy Storage Strategy Implementation Module This module reads or receives energy storage plan power strategies generated by the algorithm. The strategy table includes at least the following fields: node ID, system ID, effective date, start time, end time, power, time range, strategy type, and strategy model.

[0058] Power symbols can be used to identify actions: (1) A power of less than 0 indicates charging.

[0059] (2) A power greater than 0 indicates discharge.

[0060] (3) Power equal to 0 indicates standby.

[0061] The system generates all-day strategy points at the requested granularity, and merges consecutive charging time points into charging periods and consecutive discharging time points into discharging periods.

[0062] 2.5 Electricity Price Integration Module This module reads the node's basic time-of-use electricity price and holiday special electricity price. The basic time-of-use electricity price takes effect monthly, while the holiday special electricity price takes effect on a specific date. When both the basic electricity price and the special electricity price exist at the same time on the same date, the special electricity price overrides the basic electricity price.

[0063] For revenue calculation, the system expands the electricity price range to a unified time point and matches the corresponding electricity price at the strategy point or the actual revenue point. If the requested granularity is not the basic electricity price granularity, the system can round down to the adjacent electricity price point.

[0064] 2.6 Revenue Verification Module This module calculates actual and planned revenue using a unified financial standard.

[0065] Actual revenue is calculated using actual charging and discharging power, as well as peak, high-peak, flat, low-peak, and deep-peak electricity price fields. Actual charging cost = Σ(Price of each electricity price segment × Charging capacity of each electricity price segment) Actual discharge revenue = Σ(Price of each electricity price segment × Discharge capacity of each electricity price segment) Actual revenue = Actual discharge revenue - Actual charging cost Planned revenue is calculated using planned power output and matching electricity price: Single-point power consumption = |Planned power| × Time interval in hours Single-point amount = Single-point electricity consumption × Electricity price at the corresponding time point Planned charging cost = Σ(Amount per charging point) Planned discharge revenue = Σ(Single discharge point revenue) Planned revenue = Planned discharge revenue - Planned charging cost The daily net profit returned by the scheduling strategy is calculated as follows: Daily net income = Discharge income - Charging cost The system returns the above results after reserving them to a fixed monetary precision to avoid inconsistencies caused by recalculation by the agent or caller.

[0066] 3. Algorithm Model Section 3.1 Algorithm Overview The algorithm in this example consists of a multi-scale scene-aware load forecasting model, an uncertainty interface module, and a two-stage energy storage collaborative scheduling model with risk constraints. It is integrated with the business system's data alignment and revenue verification to form a closed loop. Load forecasting and IT load forecasting no longer use the generic PatchTST structure as the primary solution. Instead, a dedicated forecasting structure with multi-scale slicing, scene-gated fusion, critical period enhancement, and dual-head output is adopted, outputting future load point forecasts, uncertainty metrics, and critical period risk scores. Given the principal forecast values, uncertainties, electricity prices, and equipment constraints, the energy storage scheduling model first generates a baseline feasible solution that satisfies the hard constraints. Then, an improved genetic algorithm performs risk correction, revenue enhancement, and constraint repair within the neighborhood of the baseline solution, ultimately generating an executable all-day charging and discharging plan.

[0067] The point predictions, uncertainty metrics, and critical period risk scores output by the prediction model are fed into the dispatching side and used to calculate the conservative load upper limit, discharge power contraction coefficient, SOC conservative margin, and risk penalty weight, respectively, so that the prediction error can directly change the energy storage constraints and objective function.

[0068] The prediction and scheduling sides are connected through a unified risk interface: the prediction results synchronously include point prediction values, uncertainty measures, and risk scores for critical periods; the scheduling model writes the uncertainty measures and risk scores into the constraint boundaries in the first stage, and continues to use the same constraints to filter and repair candidate solutions in the second stage.

[0069] in, This represents the predicted load at a future time t. This represents a measure of prediction uncertainty. H represents the risk score for the critical period, and H represents the prediction window length. In this specification, when a symbol for t is present, t enclosed in parentheses or as a subscript indicates the value of that symbol at time t.

[0070] 3.2 Load Forecasting Model A multi-scale scenario-aware load prediction model for computing and power coordination nodes. The model input and output are uniformly defined as follows: (1) Historical power load sequence : The load points of a node within the past L sampling points in 5 or 15 minutes, where R represents the set of real numbers.

[0071] (2) IT load sequence : Server, data center or information technology equipment load point, used to characterize the impact of computing power services on the total load.

[0072] (3) Characteristics of computing load and operating status , This includes computing power utilization, data center operating status, node type, line name, and energy storage system identification.

[0073] (4) Calendar and electricity price scenario characteristics This includes labels for weekdays, weekends, holidays, hours, minutes, months, seasons, peak, flat, valley, high, low, and special electricity prices.

[0074] (5) Optional external features This includes factors such as weather, temperature, humidity, production plans, and park occupancy rates, used to supplement external driving factors.

[0075] (6) The model output includes point predictions for the next H time points. Uncertainty measure Risk score during key periods , as well as the confidence interval or conservative load upper limit available for use by the business side.

[0076] All parameters are taken from actual engineering configurations. The following descriptions are based on input / output, structure, training, and inference processes.

[0077] The node load prediction model employs multi-scale slicing, scene-gated fusion, key period enhancement, and a dual-head output structure to simultaneously handle local fluctuations, cross-period trends, and scene differences.

[0078] Model input and output. The historical window length is denoted as L, and the prediction window length is denoted as H; for each future time t, the model outputs the predicted point value. Uncertainty measure and risk score . Used to describe the predicted volatility or the width of the confidence interval at that moment. It is used to characterize scheduling risks in key scenarios such as peak-valley boundaries, peak electricity price switching, and historical error high-incidence intervals.

[0079] The model sequentially performs multi-source input alignment and normalization, short-scale and long-scale slice encoding, scene-gated fusion, attention enhancement for critical periods, and output computation. The point prediction head and uncertainty head are learnable output heads, and the risk score for critical periods is calculated from the uncertainty metric and the critical period identifier.

[0080] The specific network structure is as follows: The input embedding layer consists of a linear mapping of continuous variables and concatenation of categorical features. Continuous variables include historical load, IT load, temperature and humidity, computing power load, and electricity price. Categorical features include node type, hour, weekday, holidays, peak / flat / valley labels, and energy storage system identifiers. The short-scale branch uses one-dimensional convolution and short patch encoding to capture rapid fluctuations at a 5-minute or 15-minute granularity; the long-scale branch uses long patch encoding and Transformer encoding to capture intraday cycles and cross-period trends. The scene gating layer uses a multilayer perceptron (MLP) to generate gating weights and uses a Softmax activation function to weightedly fuse short-scale and long-scale features. The key period enhancement layer adds biases for peak / valley boundaries, peak electricity prices, and historical error high-incidence intervals to the attention score. The output layer includes a point prediction head, an uncertainty head, and a risk score calculation unit. The point prediction head and uncertainty head are learnable output heads. The risk score calculation unit calculates the risk score based on the uncertainty measure and key period identifier. All three results are written to the business field and read by the scheduling model.

[0081] In this article, t Indicates the predicted or scheduled time. x This represents the input sequence; parameters with different module subscripts are not reused.

[0082] in, For a moment Scene vectors, As scene gating weights, It is a short-scale feature. Long-scale features As a feature of fusion, For scene gating weighting coefficients, For scene gating bias. For short-scale scenarios, gating weights For long-scale scene gating weights.

[0083] in, i To query the time index, j For key-time index, For query, As key, For feature dimension, To improve the attention score, To enhance attention scores before the enhancement The 0 / 1 indicator represents peak or high-peak periods. The 0 / 1 identifier indicates the peak / valley or electricity price switching boundary. The 0 / 1 marker indicates a period with a high incidence of historical errors. , , This is the enhancement coefficient.

[0084] Multi-scale slice layer. Short-scale slices are used to characterize local rapid fluctuations. The length and sliding step size of the short-scale slice are denoted as follows: , Long-scale slices are used to depict trends across time periods. The length of the long-scale slice and the sliding step size are denoted as follows: , ,and For any input sequence The short-scale slice set is Short-scale slice starting point satisfy Long-scale slice set is Long-scale slice starting point satisfy Short-scale features were obtained from the two types of slices through linear mapping. Long-scale features .

[0085] in, and These are the starting indexes of the short-scale slice and the starting index of the long-scale slice, respectively. and These are the trainable linear mapping weight matrices for short-scale and long-scale slices, respectively. and These are the corresponding bias vectors.

[0086] The encoder features are fed into the point prediction head and uncertainty head respectively, yielding the point prediction value and positive uncertainty measure; the risk score for critical periods is calculated jointly by the normalized uncertainty measure, peak or high-peak period identifier, electricity price switching boundary identifier, and historical error high-incidence period identifier. Both the point prediction head and the uncertainty head are trainable output heads. The raw output of the uncertainty head is converted to a positive value using the softplus function. Then normalize according to the preset quantile values ​​of the training or validation set. To [0, 1]. , , and The weights are non-negative, and the sum of the four is 1.

[0087] The historical window length L is set to 288 as an engineering example of a 24-hour historical window with a 5-minute granularity; when the sampling granularity is 15 minutes, it can be converted to 96 based on the same time coverage.

[0088] Specific uses of short-scale branches and Long-scale branches are used in specific ways. and These two controls the length of the corresponding slice and the sliding step size, respectively.

[0089] Prediction step size. The prediction window length H is set to 288 as the prediction instance for 5 minutes in the next 24 hours. The model outputs at each future time step. , and This is for subsequent energy storage scheduling; the business side can extract or aggregate from the 5-minute prediction points into 96 15-minute points or 24 1-hour points according to the request granularity.

[0090] Output granularity. The algorithm side outputs a 5-minute granularity by default to preserve details of load fluctuations, peak electricity price boundaries, and data center load changes. When the business side outputs at a 15-minute or 1-hour granularity, in addition to aggregating the predicted values ​​at the aggregation points, it also aggregates uncertainty metrics and risk scores to prevent the scheduling side from losing risk information.

[0091] Loss function. Training uses joint loss. ,in, This represents the prediction error for all time points. The weighted error is calculated based on peak-valley boundaries, peak periods, and periods with a high incidence of historical errors. For the uncertainty calibration loss, To monitor losses based on risk scores, , , where is the weighting coefficient. Through this joint loss, the model optimizes both the average prediction accuracy and constrains the availability of risk outputs during critical periods.

[0092] Wherein, the prediction error is denoted as , If it is true, The predicted values ​​are used; the weights for key periods are obtained by weighting peak or high-peak period identifiers, switching boundary identifiers, and historical error high-incidence identifiers; the risk score supervision label is set to 1 when any of the above identifiers is 1. The prediction loss for all time periods uses mean squared error, the loss for key periods (peak-valley boundary, peak period, and historical error high-incidence period) uses weighted mean squared error, the uncertainty calibration loss uses Gaussian negative log-likelihood, and the risk score supervision loss uses binary cross-entropy; all weights are non-negative and determined by the validation set.

[0093] Training sample construction method. A synchronous sliding window is used to construct multi-source samples. Each sample contains historical load, IT load, computing load, electricity price, calendar and node status features of length L, and a label sequence of length H. The window continuously slides along the time axis to generate training samples, and is divided into training set, validation set and test set according to time order to avoid future information leakage; risk labels are jointly generated by peak-valley switching, peak electricity price, historical error distribution and business critical periods.

[0094] Outlier handling methods: The input side first performs timestamp alignment, outlier sampling point identification, and sample-level normalization; for device jumps, acquisition anomalies, and extreme outliers, the following methods can be used: Outliers are marked using quantile limits or business rules. These outliers are not directly used for ordinary error learning, but instead enter the risk labeling or uncertainty calibration process, enabling the model to output higher risk scores in abnormal scenarios.

[0095] Missing value handling methods: For missing points in historical sequences, forward imputation and linear interpolation are first used to ensure input continuity; when there are consecutive missing values ​​or key covariates are missing, a missing value mask is added to the sample and the uncertainty of the corresponding time period is increased. The model output, in addition to the prediction curve, is output synchronously. When the business side completes the curve, the risk field must not be discarded.

[0096] Model update frequency. The model can be retrained daily or during off-peak business windows. During updates, it's necessary to simultaneously verify point prediction error, critical period error, uncertainty calibration effectiveness, and scheduling benefit pullback indicators. Only when these indicators meet the deployment threshold should a new prediction model version be switched to, avoiding scheduling risks caused by deploying a model solely based on a decrease in MSE.

[0097] Model Principle. The core of the multi-scale scene-aware load prediction model lies in simultaneously modeling local fluctuations, cross-period trends, and differences in business scenarios, and explicitly outputting the predicted risk to the scheduling model. Its processing chain includes: multi-source input alignment, short / long-scale slicing, scene gating fusion, key period enhancement encoding, dual-head output, and joint loss training.

[0098] (1) Multi-scale slices. Short-scale slices capture rapid load fluctuations and peak boundaries, while long-scale slices capture intraday cycles, cross-period trends, and slow load change patterns. The two types of slices are modeled in parallel and then fused to avoid the fact that a single slice scale can only be biased towards one of the local or trend aspects.

[0099] (2) Scene gating. Node type, time period type, electricity price tag, data center operating status, and computing load together form a scene vector. The gating layer is based on The model adaptively adjusts the feature weights of short-scale and long-scale to enable different responses in scenarios such as weekday peak hours, holiday off-peak hours, and peak electricity price switching.

[0100] (3) Enhancement during critical periods and dual-head output. The model adds enhancement weights to the peak-valley boundary, peak periods, and historical error high-incidence intervals in the attention score, and outputs the prediction head. Output via uncertainty header Then, by combining the peak label and the switching label, we can obtain... This structure allows the prediction results to be directly used in risk constraint scheduling.

[0101] Table 1 The output of the load forecasting model is a structured field, including at least the following: and Identify the node and energy storage system to which the prediction results belong; The timestamp corresponding to the prediction point; Predicted load values ​​for points ; predict uncertainty For uncertainty measurement ; Risk score for critical periods ; This is a conservative upper limit for loads used by the dispatching side; Indicates adjustable load margin; Used to assess maximum power demand. These fields are written in a structured format to the forecast data table, which is read jointly by the scheduling model and the business revenue module.

[0102] 3.3 IT Load Forecasting Model The IT load forecasting model and the node load forecasting model share the same basic architecture design, but do not simply share all model parameters. The two types of tasks are configured with separate task identifier embeddings, input normalization parameters, task adapters, and independent output headers. In joint training scenarios, task loss weights limit the excessive pull of any task on the shared encoding layer; in separate training scenarios, only the architecture design is reused, not the weights. This avoids the biases introduced between IT load forecasting and node total load forecasting due to different distributions of target variables and covariates.

[0103] The IT load model's inputs primarily include historical IT load, computing power utilization, data center operating status, temperature, and humidity; the node load model's inputs primarily include total node load, electricity price scenarios, calendar scenarios, energy storage system status, and optional external features. Both models maintain consistent output fields, including... , and This is so that the subsequent scheduling module can read the data using a unified risk interface.

[0104] The IT load forecasting model shares a multi-scale scenario perception backbone structure with the node load forecasting model, but their input covariates and target variables differ. In addition to historical IT load, its inputs primarily include temperature, humidity, computing load, and data center operating status; its outputs are the predicted IT load for the next 24 hours, uncertainty metrics, and risk scores, named ITForecastLoad, which is subsequently read and integrated.

[0105] The IT load forecasting model follows the above-mentioned backbone structure; its differentiated configurations are explained below.

[0106] Model Name. IT load forecasting uses the same multi-scale scenario-aware forecasting backbone as node load forecasting, but the target variable is a single variable output of IT load. This configuration not only utilizes multi-source covariates to improve the accuracy of IT load forecasting, but also retains uncertainty and risk scores, facilitating the transmission of risk to the scheduling side when there is a sudden increase in data center load or anomalies in cold load.

[0107] Input window length and prediction step size. IT load forecasting also uses historical window L and prediction window H. In the engineering example, L=288 and H=288, corresponding to the previous 24 hours of history and the next 24 hours of prediction at a 5-minute granularity. The short-scale / long-scale slice parameters are consistent with the node load model and can be finely adjusted individually according to the IT load fluctuation frequency.

[0108] Output granularity. The IT load forecast defaults to outputting 5-minute point forecast values, uncertainty, and risk scores; the business side aggregates these three types of fields simultaneously when displaying them at 15-minute or 1-hour intervals, so as to calculate the mean absolute error (MAE), mean absolute percentage error (MAPE), forecast accuracy, and critical period deviation at the same timestamp.

[0109] Loss function and input features. The training adopts the joint loss consistent with node load prediction. In addition to historical IT load, the input features also include temperature, humidity, computing load, and data center operating status. Among them, computing load describes the driving effect of data center computing demand on IT load, and risk output is used to alert scenarios such as high data center load, abnormal temperature and humidity, or sudden increase in computing power.

[0110] Other hyperparameters and training procedures: The number of encoder layers, gating dimension, slice scale, risk score weights, learning rate, and early stopping strategy can inherit the default configuration of the node load prediction model; in addition to MAE / MAPE, online validation also needs to check the error and uncertainty coverage during high load periods.

[0111] When the business system does not specify a target date, it uses a scrolling window: Base time = current time rounded down to the previous 5 minutes; Actual window = 12 hours before the base time to the base time; Forecast window = 12 hours before the baseline time to 24 hours after the baseline time; Prediction bias and accuracy are calculated as follows: MAE = Σ|Actual Load - Forecast Load| / Number of Actual Points that Can Be Aligned; MAPE = Average value (|Actual load - Forecast load| / |Actual load|) × 100%; Accuracy = 100% - MAPE; When the actual value is 0, the point participates in the MAE calculation but not in the MAPE calculation, thus avoiding meaningless percentage errors caused by dividing by 0.

[0112] 3.4 Energy Storage Optimization Model The energy storage optimization model generates a full-day charge / discharge plan based on point forecasts, uncertainty metrics, risk scores, electricity prices, date types, and equipment constraints. The objective function consists of basic revenue, risk penalties, action switching penalties, and constraint repair costs.

[0113] Disclosure constraints include: (1) The charging power shall not exceed the maximum charging power of the energy storage device.

[0114] (2) The discharge power shall not exceed the maximum discharge power of the energy storage device.

[0115] (3) Charging and discharging must not be done at the same time.

[0116] (4) The remaining percentage of energy storage capacity (SOC) shall not be lower than the minimum SOC or higher than the maximum SOC.

[0117] (5) The SOC at the end of the period remains within the preset target range.

[0118] (6) Charge and discharge efficiency participates in SOC renewal.

[0119] (7) The number of charging and discharging cycles, the minimum duration, or the specified disabled period can be limited according to the business strategy.

[0120] The following device parameters use anonymized example values.

[0121] The optimization method employs a two-stage risk-constrained collaborative scheduling approach. In the first stage, a mixed-integer linear programming or equivalent benchmark scheduling model is used to obtain a benchmark feasible solution under the constraints of power boundary, SOC boundary, charge-discharge mutual exclusion, end-of-period SOC target, and discharge power and SOC safety boundary determined by predicted risk. In the second stage, an improved genetic algorithm is used to search within the neighborhood of the benchmark solution, requiring candidate solutions to continue to satisfy the constraints of the first stage.

[0122] The first stage uses point prediction values, electricity prices, equipment parameters, uncertainty metrics, and risk scores as inputs. The risk mapping results are then incorporated into the upper limit of discharge power, the lower limit of state of charge (SOC), and a risk penalty term to obtain the initial power sequence and action state sequence. This baseline solution provides a feasible search center for the second stage.

[0123] Risk mapping. The mapping relationship from predicted risk to scheduling constraints includes: in, To be a conservative upper limit of the load, The discharge power contraction coefficient, For SOC conservative margin, Risk penalty weight; , , , , , , , For mapping coefficients, This represents the minimum discharge ratio lower limit. During high-risk periods, the allowable discharge intensity is automatically reduced, the SOC safety boundary is increased, and a risk penalty is added to the objective function.

[0124] In the formula, This is the basic SOC safety margin reserved when both the uncertainty measure and the risk score are 0; max indicates that the maximum value is used for calculation. The minimum allowable value is the lower limit of the discharge power contraction coefficient. Each mapping coefficient takes a non-negative value, and boundary truncation ensures that the discharge power contraction coefficient is located in (0, 1], the SOC conservative margin does not exceed the available SOC range, and the risk penalty weight is not less than 0.

[0125] The second stage employs action segment encoding, consisting of action type, power level, and duration, and injects the baseline solution from the first stage and its neighboring perturbation solutions into the initial population. For candidate solutions... n Fitness is calculated using the following formula: Among them, the basic income item This represents the difference between discharge revenue and charging cost, and includes a risk penalty item. This indicates the penalty for exceeding the allowable discharge power during high-risk periods; action switching penalty item. Used to limit frequent charge / discharge switching and constrain repair cost items. Record the amount of adjustment required to restore the feasibility of candidate solutions. Number the candidate solutions. For the first The fitness of each candidate solution , , and These are the basic benefit term, risk penalty term, action switching penalty term, and constraint repair cost term for the candidate solution, respectively. For a moment Applicable time-of-use electricity pricing, The duration is the length of the time period. and Candidate solutions At any moment Non-negative discharge power and non-negative charging power, For a moment Risk penalty weight, The discharge power contraction coefficient, The maximum allowable discharge power of the equipment. For candidate solutions At any moment The action state is defined by I(·), which is an indicator function that takes the value 1 if the condition is met and 0 otherwise. For candidate solutions n The cost of constraint repair and adjustment at time t; , and The non-negative weights are the corresponding penalty terms.

[0126] Neighborhood disturbances are generated by changing the power level, duration, or boundary of an action segment, altering only one parameter level at a time, and constraint repair is performed after the disturbance. For high-risk periods, the discharge power disturbance amplitude of candidate solutions is limited by both the discharge power contraction coefficient and the risk score; the higher the risk, the smaller the allowable positive discharge disturbance.

[0127] The upper limit of discharge power disturbance can be expressed as: ,in k The non-negative risk adjustment coefficient and These are the minimum and maximum permissible disturbance amplitudes for the equipment, respectively. clip For amplitude limiting operation, This is the initial value of the disturbance amplitude.

[0128] In the formula, the basic revenue item is calculated based on the time-of-use electricity price, the duration of the time period, and the charging and discharging power; the risk penalty item is used to penalize the discharge power exceeding the risk contraction limit; the action switching penalty item is calculated based on the number of action changes in adjacent time periods; and the constraint repair cost item records the adjustment amount required for recovery power, SOC, charging and discharging mutual exclusion, and duration constraints.

[0129] Constraint repair is performed in the following order: First, the charging and discharging power is truncated to the device boundary; then, the simultaneous charging and discharging states during the same period are eliminated; subsequently, adjacent action segments are checked forward and adjusted back according to the SOC evolution to ensure that the SOC of each period and the final SOC satisfy the constraints; finally, action segments shorter than the minimum duration are merged. Candidate solutions that are still infeasible after repair are not included in the next generation.

[0130] Decision variables. The scheduling period is divided into T time periods, and the first stage is set... , , , and The second stage of chromosome formation consists of action segments. Power rating It consists of the duration and retains the offset from the baseline solution to limit the search to the neighborhood of feasible solutions. and They are time points The charging state variable and the discharging state variable take values ​​of 0 or 1, and not both of them take the value of 1 at the same time; and They are time points Non-negative charging power and non-negative discharging power; For a moment The state of charge; In the action segment state, The power level of the action segment is denoted by , and duration is the duration of the action segment. The offset from the reference solution is used to characterize the power level, duration, or boundary changes of the action segment relative to the reference solution.

[0131] Basic returns based on time t Applicable time-of-use electricity pricing Length of time period Non-negative discharge power Non-negative charging power Calculation; when the electricity price unit is yuan / kWh and the power unit is kW and When the unit is hourly, the revenue from discharge in a single time period minus the charging cost is: .

[0132] SOC update formula. State of charge is based on... Updated periodically; in risk-constrained scheduling, it is also necessary to meet the following requirements. This allows for additional safety margins during periods of high uncertainty. This is the minimum allowable value for the state of charge. This represents the maximum allowable value for the state of charge.

[0133] in, and These refer to the non-negative charging power and non-negative discharging power in the mathematical model, respectively. For charging efficiency, For discharge efficiency, The duration is the length of the time period. This represents the rated capacity of the energy storage. The power field in the engineering output still uses negative values ​​to represent charging, positive values ​​to represent discharging, and zero values ​​to represent standby; the mathematical model internally uses non-negative charging and discharging variables to avoid mixing of symbols.

[0134] Charge / discharge efficiency. Charging efficiency. With discharge efficiency Using a value of 0.9 (example), this means that approximately 90% of the energy is effectively retained during the charge-discharge conversion, while losses are included in SOC updates and revenue calculations.

[0135] Maximum capacity. Rated energy storage capacity. Take 2 MWh (example) as the benchmark for SOC normalization and the upper limit of available energy.

[0136] Maximum power limit. Maximum charge / discharge power. (Example); The SOC operating range is limited to 10%~90% (corresponding to approximately 0.2~1.8 MWh of usable energy), ensuring sufficient usable capacity while guaranteeing battery cycle life.

[0137] Solution process: First, feasible plans are obtained from the baseline model in the first stage; then, the baseline solution, neighborhood perturbation solutions, and a small number of random feasible solutions are combined to form an initial population; subsequently, fitness calculation, selection, crossover, mutation, constraint repair, and elite retention are performed sequentially; when the maximum number of generations is reached or fitness convergence is achieved, the final policy is output. Typical parameters are population size 100, maximum number of generations 200, crossover rate 0.8, and mutation rate 0.05 (example).

[0138] A fallback strategy for anomalies or when no feasible solution is available. When the principal value of the forecast is missing, it is filled using the daily average or the curve of the most recent working day, and... Set to high risk; relax non-critical soft constraints and retain hard constraints when the baseline solution is infeasible; revert to the rule-based strategy when the improved genetic algorithm is still infeasible after multiple restarts; output a standby strategy for all day if the expected return after risk correction is negative.

[0139] The first stage simultaneously applies upper limits to charging and discharging power, mutual exclusion between charging and discharging, upper and lower limits to State of Charge (SOC), a target SOC at the end of the period, an upper limit to discharge power after risk contraction, and a lower limit to SOC after risk increase; among which the end-of-period constraint applies to SOC(T). Candidate solutions in the second stage continue to satisfy all constraints of the first stage, and excessive switching is limited through action switching penalties and constraint repair.

[0140] In the above process, the predicted risk parameters enter the energy storage constraints from the first stage. The benchmark solution of the first stage limits the search center of the second stage. The second stage generates an executable plan through action segment encoding, risk adaptive perturbation and constraint repair.

[0141] The optimized model outputs a structured planned power record, including at least the node identifier, energy storage system identifier, effective date, start time, end time, power, strategy type, model version, risk score, SOC margin, and remediation identifier. A negative power value indicates charging, a positive value indicates discharging, and 0 indicates standby.

[0142] Specific application examples The following example illustrates the predicted scheduling of node snd1 on the target date of 2026-03-02.

[0143] I. Node Load Forecasting The system receives node load forecasting requests: { "nodeCode":"snd1", "targetDate":"2026-03-02", "granularity":"FIFTEEN_MINUTES" } The system maps snd1 to an internal node ID, uses a multi-scale scene-aware load prediction model to generate 5-minute predicted load data between 0:00 and 24:00 on the target date, and outputs it synchronously. predict uncertainty and If the requested granularity is 15 minutes, the system aggregates the 5-minute prediction points into a 96-point prediction curve in 15-minute increments, while retaining the uncertainty and risk score for the corresponding time period.

[0144] When aggregating 5-minute forecast power into 15-minute power, the business side uses an interval arithmetic mean; when calculating the corresponding electricity consumption, the power is multiplied by the time period length. The aggregation uncertainty measure uses the upper quantile value of the interval, and the aggregation risk score uses the maximum value of the interval, in order to retain conservative risk information for critical periods.

[0145] Total power consumption = Σ(predicted power × 15 / 60) Peak power = max(predicted power) Peak time = the time point corresponding to peak power The output includes node ID, target date, total power consumption, peak power, peak time, prediction curve, and report download address.

[0146] II. IT Load Forecasting When the user does not specify a target date, the system automatically constructs a scrolling window. Assuming the current time is 10:17, the base time is rounded down to 10:15. The actual curve window covers the previous 12 hours, and the predicted curve window covers the period from the previous 12 hours to the next 24 hours. The system queries the IT forecast curve, uncertainty, and risk score generated by the multi-scale scenario-aware IT load forecasting model, and calculates MAE, MAPE, and forecast accuracy by aligning them with the actual IT load curve using the same timestamp.

[0147] This embodiment can simultaneously answer two questions: "What will the IT load trend be in the next 24 hours?" and "What is the forecast deviation in the last 12 hours?"

[0148] III. Energy Storage Dispatch The system receives a scheduling policy request: { "nodeCode":"snd1", "targetDate":"2026-03-02", "granularity":"FIFTEEN_MINUTES" } The system reads the energy storage planned power table generated by the two-stage risk-constrained scheduling model and generates a 96-point strategy table for the entire day according to the target date. Before scheduling, it reads the point prediction values, uncertainty metrics, and risk scores, calculates the conservative load upper limit, discharge power contraction coefficient, SOC conservative margin, and risk penalty weight, and then performs the first-stage benchmark solution and the second-stage neighborhood optimization.

[0149] When both high uncertainty and high risk scores occur during peak electricity price periods, the system reduces the maximum allowable discharge power and increases the lower limit of SOC during that period. The candidate solutions generated in the second stage still need to meet the above constraints to avoid over-discharge caused by prediction bias.

[0150] 1. If the power is less than 0, it is identified as a charging point, and the charging cost is calculated.

[0151] 2. If the power is greater than 0, it is identified as a discharge point, and the discharge benefit is calculated.

[0152] 3. If the power is equal to 0, it is identified as a standby point and will not be included in the revenue calculation.

[0153] The system merges continuous charging points into charging periods and continuous discharging points into discharging periods, ultimately outputting total charging costs, total discharging revenue, daily net revenue, and a full-day strategy table.

[0154] IV. Energy Storage Arbitrage Profits The system receives a request for monthly revenue inquiry: { "nodeCode":"snd1", "targetPeriod":"2026-03" } The system converts March 2026 into a query window covering the period from the first day of the current month to the first day of the following month. Actual revenue is derived from the actual revenue table, while planned revenue is derived from a re-verification of planned power and electricity price records. The system returns: actualRevenue: Actual revenue actualChargeCost: Actual charging cost actualDischargeRevenue: Actual discharge revenue plannedRevenue: Planned revenue plannedChargeCost: Planned charging cost plannedDischargeRevenue: Planned Discharge Revenue This application also provides a system comprising one or more functional units or modules for performing the method described in any of the above embodiments.

[0155] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0156] The electronic device may be a server, gateway device, desktop computer, or other computing device capable of executing the computer program.

[0157] The processor may be a central processing unit, a digital signal processor, an application-specific integrated circuit, a field-programmable gate array, or other processor capable of executing the computer program.

[0158] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0159] The memory includes a program storage area and a data storage area; the program storage area stores the computer program that implements the method, and the data storage area stores node load, forecast results, energy storage plans, electricity price data, and revenue verification records. The memory may include random access memory and non-volatile memory.

[0160] This application also provides a computer-readable storage medium storing a computer program or instructions thereon; when the computer program or instructions are executed by a processor, they implement the steps of the above-described method embodiments. The computer-readable storage medium may include a read-only memory, random access memory, solid-state memory, magnetic disk, or optical disk, or other storage media with non-volatile characteristics.

[0161] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0162] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A load forecasting and energy storage scheduling method for computing-power collaborative scenarios, characterized in that, include: On the forecasting side, load forecasting data is obtained, which includes continuous variables and categorical features. The continuous variables include electricity load, IT load, and electricity price. The categorical features include node type, time period type identifier, scenario label, critical period label, and energy storage device identifier. The trained load forecasting model is used to process the load forecasting data to obtain load forecasting results. The load forecasting results include forecasted load, forecast uncertainty measure, and critical period risk score. The forecast uncertainty measure is used to describe the forecast fluctuation or confidence interval width at future times, and the critical period risk score is used to characterize the scheduling risk during critical periods. The load prediction model comprises an input embedding layer, a short-scale branch, a long-scale branch, a scene gating layer, a key period enhancement layer, and an output layer. The input embedding layer concatenates the continuous variables and categorical features to obtain concatenated features. The short-scale branch captures rapid fluctuations in the concatenated features over short time periods to obtain short-scale features. The long-scale branch captures trends in the concatenated features over long time periods to obtain long-scale features. The scene gating layer performs weighted fusion of the short-scale and long-scale features based on gating weights to obtain fused features, where the gating weights are determined based on the scene labels. The key period enhancement layer enhances the fused features using an attention mechanism to obtain enhanced features, where the attention score is defined by an enhancement coefficient determined by the key period labels. The output layer outputs the load prediction result based on the enhanced features. On the dispatching side, the load forecast results are obtained; using the constructed energy storage optimization model, a charging and discharging plan for the energy storage equipment is generated based on the load forecast results and the electricity price; wherein, in the first stage, the forecast uncertainty metric and the critical period risk score are mapped to discharge power constraints, state of charge constraints and risk penalty weights, and a baseline solution that satisfies the constraints is generated by combining the electricity price and energy storage equipment parameters; in the second stage, an improved genetic algorithm is used to search in the neighborhood of the baseline solution to make the candidate solutions continue to satisfy the constraints, and the final solution is obtained.

2. The method as described in claim 1, characterized in that, Key time periods include: peak or high-peak periods, peak-valley or electricity price switching boundaries, and / or periods with a high incidence of historical errors; the risk score of the key time period is defined by the prediction uncertainty measure, peak or high-peak period identifier, peak-valley or electricity price switching boundary identifier, and / or period with a high incidence of historical errors identifier.

3. The method as described in claim 1, characterized in that, The short-scale branch uses one-dimensional convolution and short data block encoding, while the long-scale branch uses long data block encoding and Transformer encoding.

4. The method as described in claim 2, characterized in that, To query the time index, Key moment index; basic attention score Enhancement items during key periods Revised to The key period enhancement is determined by a weighted sum of peak or high-peak period identifiers, peak-valley or electricity price switching boundary identifiers, and historical error-prone period identifiers: in, This is an indicator for peak or high-peak hours. This serves as a boundary marker for peak / valley or electricity price switching. This serves as a marker for periods with a high incidence of historical errors. , and For weighted weights.

5. The method as described in claim 1, characterized in that, The loss function of the load forecasting model for: in, This represents the prediction error for all time points. This is the weighted error for the critical period. For uncertainty calibration items, To monitor losses based on risk scores, , , For weights.

6. The method as described in claim 1, characterized in that the constraint Including: upper limit of conservative load Discharge power contraction coefficient SOC conservative margin and risk penalty weight ,in: in, For a moment Point-based load forecasting For a moment The measure of prediction uncertainty For a moment Risk score during key periods , , , , , , , For mapping coefficients, This is the minimum discharge ratio. It is the basic SOC safety margin reserved when both the uncertainty measure and the risk score are 0, and max means to calculate the maximum value.

7. The method as described in claim 6, characterized in that, An improved genetic algorithm is used to optimize the baseline solution, specifically including: Set up a chromosome and initialize a population. The chromosome is set as an action segment code consisting of action type, power level and duration. The initial population includes: the benchmark solution and the neighborhood perturbation solution of the benchmark solution. The fitness calculation, selection, crossover, mutation, constraint repair, and elite retention processes are executed sequentially until convergence.

8. The method as described in claim 7, characterized in that, The fitness F Expressed as: In the formula, the basic income item This represents the difference between discharge revenue and charging cost, and includes a risk penalty item. This indicates the penalty for exceeding the allowable discharge power during high-risk periods; action switching penalty item. Used to limit frequent charge / discharge switching and constrain repair cost items. Record the adjustment amount required to restore the feasibility of candidate solutions; , and The weights are non-negative.

9. A load forecasting and energy storage dispatching system for computing-power collaborative scenarios, characterized in that, It includes one or more functional units or modules for performing the method of any one of claims 1-8.

10. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-8.

Images