An energy storage demand control method based on prediction confidence self-adaption

Abstract

The application discloses a kind of based on prediction confidence self-adaptive energy storage demand control method comprising the following steps: preparation stage, scene identification and configuration selection are carried out, current date type is identified, and corresponding configuration parameters are loaded according to the identification result;Execution stage, the closed-loop control mechanism of rolling execution is used, the closed-loop process of "prediction-evaluation-adjustment-optimization-execution-feedback" is completely executed every first time period, and real-time accurate regulation and control to energy storage system is realized by dynamic iteration;Learning stage, when the execution stage is completed, the actual operation data of the whole cycle of the day is used to incrementally update the prediction model.The application solves the technical problems of the existing energy storage demand control method due to large load prediction error, lack of robustness to prediction error, unreasonable safety margin strategy and lack of multi-scenario adaptability.

Description

Technical Field

[0001] This invention relates to the field of power system energy storage control technology, and in particular to an adaptive energy storage demand control method based on prediction confidence. Background Technology

[0002] With the advancement of electricity market reform and the widespread application of peak-valley pricing mechanisms, industrial and commercial users face the challenge of high demand-side electricity costs. According to current electricity pricing policies, demand-side electricity prices are calculated based on the maximum demand within a billing cycle (e.g., monthly), and this cost can account for 30% to 50% of total electricity bills. Therefore, deploying energy storage systems to discharge during peak hours to reduce demand has become an effective way to reduce electricity costs. However, the core challenge of energy storage optimization control lies in the inevitable errors in load forecasting. The forecasting process is often affected by various factors, including weather changes (temperature, humidity), production plan fluctuations, equipment start-up and shutdown status, and unforeseen events. The error range is typically between 5% and 30%. When the forecast error is large, optimization control strategies relying on forecast data are prone to failure, causing the energy storage system to fail to achieve the expected demand management effect, and may even have negative consequences.

[0003] Currently, the main methods for controlling energy storage demand include the following categories: (1) Average discharge strategy, which distributes the energy storage capacity evenly over the control period and discharges it at a constant power. It is simple and reliable to operate and does not rely on prediction data. However, it cannot be dynamically adjusted according to actual load fluctuations, and the control effect is limited. (2) Rule-based heuristic method, which controls according to rules set by experience. It has good real-time response capability, but the control effect is unstable due to the lack of a global optimization perspective. (3) Model predictive control method, which mainly relies on load prediction and energy storage constraints to establish an optimization problem and solve the optimal discharge strategy. Typical solutions include the MPC method using a convex optimization solver. This type of method performs well when the prediction is accurate, but it is sensitive to prediction errors and has high computational complexity and poor real-time performance. Therefore, these problems manifest themselves in the following ways: First, they lack robustness to prediction errors. Traditional MPC methods rely on accurate load forecasting. When prediction errors are large, two extreme situations can easily occur: either energy storage over-discharges in the early part of the period, leading to depletion of energy storage in the middle and later stages, resulting in increased peak demand due to the lack of stored energy, or energy storage is overly conservative, leaving a large amount of unused energy storage capacity at the end of the period, resulting in low utilization. Second, the safety margin strategy is unreasonable. To cope with prediction errors, existing methods usually adopt a fixed safety margin (such as a uniform upward adjustment of 15%). However, this fixed value is too conservative and wasteful of energy storage when the prediction is accurate, and may still be insufficient when the prediction is inaccurate. Third, the optimization algorithm has high complexity. The computational complexity of MPC methods based on convex optimization solvers is typically O(n log n). First, a single optimization takes tens of seconds, resulting in poor real-time performance and high hardware requirements. Second, it lacks adaptability to multiple scenarios. The load characteristics of weekdays, Saturdays, Sundays, holidays, and the days before and after them are significantly different, but existing methods use a uniform time period configuration, which cannot adapt to different scenarios. Third, it lacks online learning capabilities. The prediction model parameters are fixed and cannot be updated according to the latest data, resulting in a gradual decrease in prediction accuracy over time.

[0004] Chinese patent application CN202311363456.X discloses a method and system for energy storage operation control based on peak-valley arbitrage. This control method operates based on an energy storage operation control system, which includes a power grid, energy storage devices, user distribution cabinets, and an operation control module. The energy storage devices consist of multiple energy storage cabinets, which are energy storage equipment. The power grid transmits electricity generated by power plants to user terminals. The energy storage devices store excess electricity during periods of low grid load, releasing the stored energy during peak load periods to balance grid load fluctuations. The user distribution cabinets distribute electricity to various electrical devices through branch circuits. The operation control module controls the operation of the energy storage devices according to user load demands and electricity consumption strategies. To address the above technical problems, this invention proposes an adaptive energy storage demand control method based on predictive confidence. Summary of the Invention

[0005] The main objective of this invention is to propose an adaptive energy storage demand control method based on prediction confidence, which aims to solve the technical problems of existing energy storage demand control methods, such as large load prediction errors, lack of robustness to prediction errors, unreasonable safety boundary strategies, and lack of adaptability to multiple scenarios.

[0006] To achieve the above objectives, the present invention provides an adaptive energy storage demand control method based on prediction confidence, wherein the adaptive energy storage demand control method based on prediction confidence includes the following steps:

[0007] S1. Preparation stage: Scene recognition and configuration selection are performed, the current date type is identified, and the corresponding configuration parameters are loaded according to the recognition results;

[0008] S2, Execution Phase: A rolling execution closed-loop control mechanism is adopted, and the closed-loop process of "prediction-evaluation-adjustment-optimization-execution-feedback" is fully executed in each first time period, so as to achieve real-time and precise control of the energy storage system through dynamic iteration;

[0009] S3. Learning phase: After the execution phase is completed, the prediction model is incrementally updated using the actual operating data of the entire cycle on that day.

[0010] One preferred embodiment is that the scene recognition in step S1 adopts a three-level priority determination mechanism, specifically as follows:

[0011] Check the match between the current date and the preset holiday list;

[0012] If a match is found, the determination bit is of type "Sunday / Holiday";

[0013] If no match is found, the weekday code is checked. If the weekday code is 6, it is determined to be of type "Saturday"; otherwise, it is determined to be of type "regular weekday" by default.

[0014] One preferred embodiment is that the configuration parameters in step S1 are specifically as follows:

[0015] Based on the identification results, the corresponding date type configuration parameters are automatically queried and loaded from the configuration parameter library. The configuration parameters include the decision start time, demand control period, control step size, and prediction reference historical data range.

[0016] One preferred embodiment is that the prediction in the closed-loop process of step S2 specifically includes:

[0017] S211. Extract historical data from the database and input it into the prediction model; the prediction model adopts a linear regression model, neural network, or deep learning model for similar days;

[0018] S212. The prediction model outputs the energy consumption sequence from the current time to the end time of demand control, and converts the energy consumption sequence into an average power sequence; the average power sequence is:

[0019]

[0020] in, for The average power consumption corresponding to the first time step of the time period. for The electricity consumption of the first time period step.

[0021] One preferred approach is to evaluate the process in step S2 within the closed-loop flow, specifically as follows:

[0022] S221. Construct a prediction error history sequence of length L, store prediction error data for L time steps, and calculate the prediction error; the prediction error is:

[0023]

[0024] in, For prediction error, For the prediction error history sequence The measured electricity consumption within a consecutive first time period, where the time step of each first time period corresponds to the end time. To predict electricity consumption;

[0025] S222. Add the prediction error to the end of the prediction error history sequence, and determine whether the queue length exceeds L. If it does, delete the first element of the prediction error history sequence.

[0026] S223. Based on the historical sequence of prediction errors, calculate... The average error at each time step and The average error at each time step respectively through average error and average error Conduct confidence assessments and dynamically adjust safety margins.

[0027] One preferred embodiment is that the adjustments in the closed-loop process of step S2 include:

[0028] S231, through average error The confidence level assessment is conducted as follows:

[0029] when When the value is ≤0.05, the confidence level is 1.0;

[0030] When 0.05 < When the value is ≤0.10, the confidence level is 0.8;

[0031] When 0.10 < When the value is ≤0.15, the confidence level is 0.6;

[0032] when When the value is greater than 0.15, the confidence level is 0.3;

[0033] S232, through average error The safety margin is dynamically adjusted, specifically as follows:

[0034] when When the value is greater than 0.10, the prediction performance is considered to have deteriorated, and the safety margin is increased by 10%; specifically: ,in, For safety margin, This is the initial value for the safety margin. This represents the upper limit of the safety margin.

[0035] when When the value is ≤0.05, the prediction performance is considered excellent, and the safety margin is reduced by 5%, specifically: , This is the lower limit of the safety margin;

[0036] When 0.05 < When the value is less than or equal to 1, the prediction performance is considered stable and the safety margin remains unchanged.

[0037] S233. The adjusted safety margin is applied to the load-average power correction, specifically as follows:

[0038]

[0039] in, For the first Predicted average power after time step adjustment For the first The original predicted average power at each time step.

[0040] One preferred solution is to optimize the closed-loop process of step S2, specifically as follows:

[0041] S241. Construct an average discharge power baseline; specifically:

[0042] Distribute the current remaining energy storage evenly to the remaining... At each time step, the average discharge power baseline is calculated, and the discharge plan is initialized with the average discharge power to form a baseline scheme.

[0043] The average discharge power baseline is:

[0044]

[0045] in, The average discharge power, This represents the current energy storage status. The remaining time steps. The time step interval;

[0046] S242, Quantifying load intensity; specifically:

[0047] Calculate the statistical characteristics of the load forecast sequence, including:

[0048] The average load is:

[0049]

[0050] The standard deviation of the load is:

[0051]

[0052] in, For average load, The standard deviation of the load;

[0053] Calculate the first based on the average load and load standard deviation. The load intensity of the time step; the load intensity is:

[0054]

[0055] in, For the first The load intensity at each time step;

[0056] S243. Confidence-weighted load intensity grading fine-tuning; specifically:

[0057] like If the value is ≤0.6, it indicates insufficient prediction reliability, and no fine-tuning will be performed. ,in, To achieve the optimal discharge power sequence, As a baseline scheme, Confidence level;

[0058] like A value greater than 0.6 indicates reliable prediction. For each time step, a graded response fine-tuning is performed based on load intensity, while a confidence-weighted control of the adjustment range is introduced; specifically:

[0059] High load : The maximum discharge increase is 25%;

[0060] Medium and high load :

[0061] The maximum discharge increase is 18%;

[0062] Mild high load :

[0063] The maximum discharge increase is 10%;

[0064] Normal load : Maintain the baseline;

[0065] low load :

[0066] The maximum discharge reduction is 25%.

[0067] S244, Smoothing constraint processing; specifically:

[0068] Set the maximum allowed change. A gradual adjustment strategy is adopted to achieve smooth constraints, specifically: the optimal discharge power sequence at the first time step. Keep it unchanged, every subsequent time step ,like Then let ,in, For the updated time step The optimal discharge power sequence; if Then let Otherwise, remain unchanged;

[0069] S245, Discharge Total Constraint Calibration; specifically:

[0070] The planned total discharge after smoothing constraints is calculated, and the total discharge is calibrated by uniformly scaling all time steps using a scaling factor; the planned total discharge is:

[0071]

[0072] The scaling factor is:

[0073]

[0074] The total discharge amount is:

[0075]

[0076] in, For the planned total discharge, Total discharge amount For time step interval, This represents the current energy storage status. This is the scaling factor;

[0077] S246. Discharge power upper and lower limit constraints; the sequence of discharge power is as follows:

[0078]

[0079] in, It is a sequence of discharge power. For energy storage discharge power, [Minimum discharge power of energy storage, maximum discharge power of energy storage].

[0080] One preferred embodiment is that the execution and feedback in step S2 of the closed-loop process are as follows:

[0081] S251, Discharge command executed, extract the first element of the discharge power sequence. As the discharge power command for the current time step, it is sent to the power conversion system of the energy storage system to control the energy storage devices according to... The discharge operation is performed, and the actual discharge capacity is calculated after the first time period; the actual discharge capacity is:

[0082]

[0083] in, This represents the actual discharge capacity.

[0084] S252, Status Update and Feedback Learning;

[0085] After the discharge operation is performed, the system updates the energy storage status. :

[0086] The max function ensures the state is non-negative, and the result is compared with the detection value of the energy storage system. Compare them; if they are inconsistent, then let... At the same time, ensure ,in, The rated capacity of the energy storage system, These are the lower and upper limits of the preset state of charge for the energy storage system, respectively.

[0087] If the system has already obtained the actual load data for the current first time period Then calculate the actual demand; the actual demand is:

[0088]

[0089] in, For actual demand, The first in the sequence of discharge power One element;

[0090] Based on actual electricity consumption Calculate real-time error Update the historical prediction error series and recalculate the average error. and average error And based on the average error and average error Update the confidence level and safety margin to provide updated parameter support for optimization at the next time step;

[0091] S253. The system clock advances by one time step and determines whether the predicted end time has been reached. If not, return to step S2 to repeat the closed-loop process. If the time has been reached, execute step S3.

[0092] In one preferred embodiment, prior to step S3, the method further includes:

[0093] Set dual trigger conditions; specifically:

[0094] Time condition verification: Determine whether the current time is greater than or equal to the end time of the prediction period;

[0095] Data integrity verification: If the time condition is met, further determine whether there are actual load data records for the end of the forecast period and the entire cycle of the day;

[0096] If the dual triggering conditions are met, incremental training is triggered; if the time condition is not met, the system remains in standby mode; if the data integrity condition is not met, the fault tolerance mechanism is activated.

[0097] One preferred embodiment is that the incremental training in step S3 employs a rolling update strategy with a fixed window length, specifically as follows:

[0098] S31. Data Preprocessing and Standardization: Extract complete daily operational data from the database and perform data preprocessing and standardization procedures.

[0099] Outlier detection, using 3 The criteria identify out-of-range abnormal data and determine the validity of the data;

[0100] Missing values ​​are filled using linear interpolation for data with a missing rate of 5% or less; data with a missing rate greater than 5% are discarded.

[0101] Feature normalization maps load data and associated features to interval;

[0102] S32. Rolling update of the training dataset: Add the preprocessed data of the day to the prediction dataset and check the size of the prediction dataset.

[0103] S33. Incremental training of the model: retrain the prediction model using the prediction dataset.

[0104] In the above technical solution of the present invention, the energy storage demand control method based on predictive confidence adaptive control includes the following steps: Preparation stage: scenario identification and configuration selection are performed, the current date type is identified, and corresponding configuration parameters are loaded according to the identification results; Execution stage: a rolling execution closed-loop control mechanism is adopted, and the closed-loop process of "prediction-evaluation-adjustment-optimization-execution-feedback" is fully executed in each first time period, achieving real-time and precise control of the energy storage system through dynamic iteration; Learning stage: after the execution stage is completed, the prediction model is incrementally updated using the actual operating data of the entire cycle of the day. The present invention solves the technical problems of existing energy storage demand control methods, such as large load prediction errors, lack of robustness to prediction errors, unreasonable safety boundary strategies, and lack of adaptability to multiple scenarios.

[0105] In this invention, a hybrid strategy that integrates average discharge robustness and model predictive control optimization capability is adopted. The average discharge power is used as a robust baseline to ensure that the energy storage balance consumption will not be exhausted prematurely. At the same time, the rolling optimization and predictive feedback mechanism of MPC are used to achieve dynamic adjustment. A regular fine-tuning method based on confidence weighting is adopted. The decision to enable optimization adjustment is made by evaluating the prediction confidence (enable if confidence is ≥0.6, maintain the baseline if <0.6), thus achieving an adaptive balance between robustness and optimization performance of the algorithm.

[0106] In this invention, a quantitative evaluation mechanism for prediction confidence is employed. By maintaining a historical queue of prediction errors over several recent time steps and calculating the average error, the confidence level is divided into four levels. The aggressiveness of the control strategy is dynamically adjusted based on the confidence level, achieving an adaptive balance of "aggressive when predictions are accurate, conservative when predictions are uncertain." Simultaneously, an adaptive dynamic adjustment mechanism for the safety margin is established. Based on the long-term average error within the range of 5-30%, the safety margin is dynamically adjusted: reducing the margin to improve efficiency when predictions are accurate, and increasing the margin to enhance safety when predictions are inaccurate. Furthermore, based on a load intensity-based graded response, all fine-tuning amplitudes are weighted by confidence levels, resulting in a computational complexity of O(n log n). It involves only simple arithmetic operations, and a single optimization takes only 15-25 milliseconds, which is faster than the traditional MPC method. It can be increased by more than 200 times.

[0107] This invention supports multi-scenario adaptive configuration, automatically adjusting time period configuration and prediction parameters according to different date types such as weekdays, Saturdays, Sundays, and holidays; it provides two optimization modes: global planning and rolling, and has an incremental training trigger mechanism to achieve automatic updates of the prediction model. Attached Figure Description

[0108] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0109] Figure 1 This is a first schematic diagram of an energy storage demand control method based on predictive confidence adaptive according to an embodiment of the present invention;

[0110] Figure 2 This is a second schematic diagram of an adaptive energy storage demand control method based on prediction confidence, according to an embodiment of the present invention.

[0111] The realization of the objective, functional characteristics and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0112] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0113] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0114] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0115] See Figures 1-2 According to one aspect of the present invention, the present invention provides an energy storage demand control method based on predictive confidence, wherein the energy storage demand control method based on predictive confidence includes the following steps:

[0116] S1. Preparation stage: Scene recognition and configuration selection are performed, the current date type is identified, and the corresponding configuration parameters are loaded according to the recognition results;

[0117] S2, Execution Phase: A rolling execution closed-loop control mechanism is adopted, and the closed-loop process of "prediction-evaluation-adjustment-optimization-execution-feedback" is fully executed in each first time period, so as to achieve real-time and precise control of the energy storage system through dynamic iteration;

[0118] S3. Learning phase: After the execution phase is completed, the prediction model is incrementally updated using the actual operating data of the entire cycle on that day.

[0119] Specifically, in this embodiment, the scene recognition step S1 adopts a three-level priority determination mechanism, which is as follows:

[0120] Holiday determination: Check the match between the current date and the preset holiday list;

[0121] If a match is found, the judgment bit is of type "Sunday / Holiday". This priority judgment result can override the natural day of the week attribute of the date.

[0122] Special workday (Saturday) determination: If no match is found, the weekday code is checked. If the weekday code is 6, it is determined to be of type "Saturday".

[0123] Determination of a regular workday: If the first two determination criteria are not met, the day will be determined as a "regular workday" by default.

[0124] This judgment mechanism, through priority nesting logic, can effectively adapt to scenarios with conflicting date attributes. For example, it can accurately classify "Sundays with make-up work" as weekdays and accurately identify "Saturdays with make-up work" as holidays, ultimately achieving the goal of full scenario coverage.

[0125] Specifically, in this embodiment, step S1, configuring parameters, specifically includes:

[0126] Based on the identification results, the configuration parameters of the corresponding date type are automatically queried and loaded from the configuration parameter library. The configuration parameters include the decision start time, demand control period, control step size, and prediction reference historical data range. See Table 1 for example parameter configuration. This differentiated configuration is based on the analysis of actual load characteristics. There is usually continuous production activity and high base load on weekday evenings. Although there is continuous high load on Saturdays, the demand control time is short. The overall load on Sundays / holidays is low and no additional demand control is required.

[0127] Table 1 Examples of parameter configuration

[0128] Date type ordinary working day Special working day (Saturday) Sunday / Holiday Decision-making start time 18:00 19:00 - Demand control period 18:05-22:00 (4 hours) 19:05-21:00 (2 hours) - Control step size 48 24 - Prediction reference historical data range 13:00-17:55 (5 hours) 14:00-18:55 (5 hours) -

[0129] Specifically, in this embodiment, the first time period is 5 minutes. This invention does not impose a specific limitation, and the specific time period can be set as needed.

[0130] Specifically, in this embodiment, after the parameter configuration is successfully loaded, the corresponding parameter configuration is automatically pushed to each core functional module based on the standardized data interface box, realizing the collaborative adaptation and linkage between modules; for example: the prediction module: using "historical data range" and "control step size" as constraints, dynamically adjusts the input feature dimension and the output prediction sequence length to ensure the consistency between the prediction results and the actual monitoring needs; the optimization module: based on "demand control period" and "control step size", constructs an optimization variable space that matches the period characteristics to achieve a balanced optimization of calculation accuracy and resource consumption; the execution module: based on the basic time unit (5 minutes) parameter, accurately controls the generation and update frequency of discharge commands to ensure the timeliness and accuracy of the control actions; adjusts the input and output according to the prediction step number and historical data range parameters; the optimization module sets the optimization variable dimension according to the period length and step number parameters; the execution module controls the update frequency of discharge commands according to the time interval parameter; the entire parameter transmission and application process is fully automated without manual intervention, providing a reliable guarantee for the seamless switching and stable operation of the system in different date scenarios; the above modules can be set according to actual needs, and this invention will not elaborate on them.

[0131] Specifically, in this embodiment, the execution phase is the core part of the entire control process. It adopts a rolling execution closed-loop control mechanism. Within each 5-minute sub-control cycle, the closed-loop process of prediction-evaluation-adjustment-optimization-execution-feedback is fully executed, and real-time and precise control of the energy storage system is achieved through dynamic iteration.

[0132] Specifically, in this embodiment, load pre-prediction is the initial step in the execution phase, providing basic data support for subsequent demand control optimization. The prediction in the closed-loop process of step S2 specifically includes:

[0133] S211. Extract historical data from the database and input it into the prediction model; the prediction model adopts a linear regression model, neural network, or deep learning model for similar days;

[0134] S212. The prediction model outputs the energy consumption sequence from the current time to the end time of demand control, and converts the energy consumption sequence into an average power sequence; the average power sequence is:

[0135]

[0136] in, for The average power consumption corresponding to the first time step of the time period. for The electricity consumption of the first time period step.

[0137] Specifically, in this embodiment, the evaluation in the closed-loop process of step S2 is as follows:

[0138] S221. Construct a prediction error history sequence of length L, store prediction error data for L time steps, and calculate the prediction error; the length L is 10, that is, the prediction error history sequence is used to store prediction error data for the most recent 10 time steps; the prediction error adopts a dimensionless relative error index, and the prediction error is:

[0139]

[0140] in, For prediction error, For the prediction error history sequence The measured electricity consumption within a consecutive first time period, where the time step of each first time period corresponds to the end time. To predict electricity consumption, the denominator uses the max function to effectively avoid division by zero when the predicted electricity consumption approaches 0; the prediction error typically ranges from 0 to 0.3, directly reflecting the degree of prediction deviation.

[0141] S222. Add the prediction error to the end of the prediction error history sequence and determine whether the queue length exceeds L. If it does, delete the first element of the prediction error history sequence and keep the clock sequence's sliding window specific.

[0142] S223. Based on the historical sequence of prediction errors, calculate... The average error at each time step and The average error at each time step respectively through average error and average error Confidence assessment and dynamic adjustment of safety margin are performed; in this invention It is 3. The value is 2. This invention does not impose specific limitations; the value can be set according to needs. Average error Used for confidence assessment, it calculates the average of the errors from the most recent three steps, focusing on fluctuations in near-term prediction performance; average error For safety margin adjustment, the average of all 10-step errors in the queue is calculated, reflecting the prediction accuracy trend over a longer period; the confidence assessment is based on the average error. A segmented mapping rule is adopted to achieve quantitative classification of prediction reliability.

[0143] Specifically, in this embodiment, the adjustments in the closed-loop process of step S2 include:

[0144] S231, through average error The confidence level assessment is conducted as follows:

[0145] when When the value is ≤0.05, the confidence level is 1.0;

[0146] When 0.05 < When the value is ≤0.10, the confidence level is 0.8;

[0147] When 0.10 < When the value is ≤0.15, the confidence level is 0.6;

[0148] when When the value is greater than 0.15, the confidence level is 0.3;

[0149] S232, through average error Dynamically adjust the safety margin to construct a buffer mechanism for prediction bias; specifically:

[0150] when When the value is greater than 0.10, the prediction performance is considered to have deteriorated, and the safety margin is increased by 10%; specifically: ,in, For safety margin, The initial value of the safety margin is 0.08. This value is a conservative parameter verified by historical operating data, and an asymmetric adjustment strategy is adopted to prioritize the safety of system operation. This is the upper limit of the safety margin, which is 0.30, to avoid resource waste caused by an excessively high safety margin.

[0151] when When the value is ≤0.05, the prediction performance is considered excellent, and the safety margin is reduced by 5%, specifically: , This serves as the lower limit of the safety margin, which is 0.05, used to ensure a basic safety buffer.

[0152] When 0.05 < When the value is less than or equal to 1, the prediction performance is considered stable and the safety margin remains unchanged.

[0153] S233. The adjusted safety margin is applied to the load-average power correction, specifically as follows:

[0154]

[0155] in, For the first Predicted average power after time step adjustment For the first The original predicted average power at each time step is corrected to reserve buffer space for prediction errors, thereby improving the robustness of system operation.

[0156] Specifically, in this embodiment, robust optimization is the core step in the execution phase. Unlike traditional MPC methods that construct convex optimization problems and call solvers like CVXPY, this invention uses a regularized robust optimization algorithm based on average discharge improvement. It uses average discharge power as a robust baseline and dynamically enables intelligent fine-tuning based on prediction confidence, achieving a balance between optimization effectiveness and computational efficiency. The algorithm's input parameters include: the current energy storage state. Remaining time steps Load forecast sequence Prediction confidence and safety margin The output is the optimal discharge power sequence. The advantage of this algorithm lies in its ability to degenerate into a robust average discharge mode at low confidence levels, while achieving precise fine-tuning based on the baseline at high confidence levels, with a computational complexity of only [missing information]. No iterative solution is required; the optimization in the closed-loop process of step S2 is as follows:

[0157] S241. Construct an average discharge power baseline; specifically:

[0158] Distribute the current remaining energy storage evenly to the remaining... At each time step, the average discharge power baseline is calculated, and the discharge plan is initialized with the average discharge power to form a baseline scheme.

[0159] The average discharge power baseline is:

[0160]

[0161] in, The average discharge power, This represents the current energy storage status. The remaining time steps. The time step interval;

[0162] S242, Quantifying load intensity; specifically:

[0163] To characterize the relative load level at each time step, the statistical characteristics of the load forecast sequence are calculated, including:

[0164] The average load is:

[0165]

[0166] The standard deviation of the load is:

[0167]

[0168] in, For average load, The standard deviation of the load;

[0169] Calculate the first based on the average load and load standard deviation. The load intensity of the time step; the load intensity is:

[0170]

[0171] in, For the first The load intensity at the time step is a dimensionless index. A value greater than 0 indicates a load above the average, while a value greater than 0 indicates a load below the average. The larger the absolute value, the more significant the load deviation.

[0172] S243. Confidence-weighted load intensity grading fine-tuning; specifically:

[0173] like If the value is ≤0.6, it indicates insufficient prediction reliability, and no fine-tuning will be performed. ,in, For the optimal discharge power sequence, As a baseline scheme, Confidence level;

[0174] like A value greater than 0.6 indicates reliable prediction. For each time step, a graded response fine-tuning is performed based on load intensity, while a confidence-weighted control of the adjustment range is introduced; specifically:

[0175] High load : The maximum discharge increase is 25%;

[0176] Medium and high load :

[0177] The maximum discharge increase is 18%;

[0178] Mild high load :

[0179] The maximum discharge increase is 10%;

[0180] Normal load : Maintain the baseline;

[0181] low load :

[0182] The maximum discharge reduction is 25%.

[0183] All of the above adjustments include confidence levels, achieving confidence weighting and ensuring that the range of fine-tuning is limited when the prediction is not sufficiently certain;

[0184] S244, Smoothing constraint processing; specifically:

[0185] To avoid drastic fluctuations in discharge power from impacting the equipment, a maximum allowable variation is set. The power variation between adjacent time steps does not exceed 30% of the baseline. A gradual adjustment strategy is adopted to achieve smooth constraints, specifically: the optimal discharge power sequence of the first time step. Keep it unchanged, every subsequent time step ,like Then let ,in, For the updated time step The optimal discharge power sequence; if Then let Otherwise, it remains unchanged; through the above-mentioned gradual adjustment mechanism, the change in discharge power is gradual, avoiding drastic jumps;

[0186] S245, Discharge Total Constraint Calibration; specifically:

[0187] The planned total discharge after smoothing constraints is calculated, and the total discharge is calibrated by uniformly scaling all time steps using a scaling factor; the planned total discharge is:

[0188]

[0189] The scaling factor is:

[0190]

[0191] The total discharge amount is:

[0192]

[0193] in, For the planned total discharge, Total discharge amount For time step interval, This represents the current energy storage status. This is the scaling factor;

[0194] S246. Discharge power upper and lower limit constraints; the sequence of discharge power is as follows:

[0195]

[0196] in, It is a sequence of discharge power. For energy storage discharge power, [Minimum discharge power of energy storage, maximum discharge power of energy storage].

[0197] Specifically, in this embodiment, the execution and feedback in the closed-loop process of step S2 are as follows:

[0198] S251, Discharge command executed, extract the first element of the discharge power sequence. As the discharge power command for the current time step, it is sent to the power conversion system of the energy storage system to control the energy storage devices according to... The discharge operation is performed, and the actual discharge capacity is calculated after the first time period; the actual discharge capacity is:

[0199]

[0200] in, This represents the actual discharge capacity.

[0201] S252, Status Update and Feedback Learning;

[0202] After the discharge operation is performed, the system updates the energy storage status. :

[0203] The max function ensures the state is non-negative, and the result is compared with the detection value of the energy storage system. Compare them; if they are inconsistent, then let... At the same time, ensure ,in, The rated capacity of the energy storage system, These are the lower and upper limits of the preset state of charge for the energy storage system, respectively.

[0204] If the system has already obtained the actual load data for the current first time period Then calculate the actual demand; the actual demand is:

[0205]

[0206] in, For actual demand, The first in the sequence of discharge power One element;

[0207] Based on actual electricity consumption Calculate real-time error Update the historical prediction error series and recalculate the average error. and average error And based on the average error and average error Update the confidence level and safety margin to provide updated parameter support for optimization at the next time step;

[0208] S253, The system clock advances by one time step, that is: The time interval is increased by 5 minutes from the previous decision time, and it is determined whether the end time of the prediction cycle has been reached. If not, the process returns to step S2 to repeat the closed-loop process. If the end time has been reached, step S3 is executed. This rolling iteration mechanism enables the system to respond to load fluctuations and prediction deviations in real time, thereby achieving dynamic adaptive control.

[0209] Specifically, in this embodiment, the process automatically starts after the execution phase is completed, using the actual operating data of the entire day's cycle to incrementally update the prediction model, enabling the model to learn the load change pattern in real time and maintain long-term prediction accuracy; before step S3, the process also includes:

[0210] Set dual trigger conditions; specifically:

[0211] Time condition verification: Determine whether the current time is greater than or equal to the end time of the prediction period;

[0212] Data integrity verification: If the time condition is met, further determine whether there are actual load data records for the end of the forecast period and the entire cycle of the day;

[0213] If the dual triggering conditions are met, incremental training is triggered; if the time condition is not met, the system remains in standby mode; if the data integrity condition is not met, the fault tolerance mechanism is activated.

[0214] The learning phase must be initiated under two triggering conditions to ensure the validity of the training data and the accuracy of the process initiation.

[0215] Specifically, in this embodiment, the incremental training in step S3 adopts a rolling update strategy with a fixed window length, specifically as follows:

[0216] S31. Data Preprocessing and Standardization: Extract complete daily operational data from the database and perform data preprocessing and standardization procedures.

[0217] Outlier detection, using 3 Criteria identify out-of-scope For abnormal data, determine the validity of the data by combining the device operation logs, and mark invalid data as missing values;

[0218] Missing values ​​are imputed using linear interpolation for data with a missing rate of 5% or less; if the missing rate is greater than 5%, the data for that day is excluded from this training.

[0219] Feature normalization maps load data and related features (such as ambient temperature, production shifts, etc.) to... Intervals are used to eliminate the impact of dimensional differences on model training.

[0220] S32. Rolling update of training dataset: Add the preprocessed data of the current day to the prediction dataset and check the size of the prediction dataset. Specifically, after adding the preprocessed data of the current day to the prediction dataset, check the size of the dataset. If the number of data days is greater than 21 days, delete the data of the oldest day by sorting by timestamp to ensure that the dataset clock maintains a fixed window of 21 days and achieves a dynamic balance between addition and elimination.

[0221] S33. Incremental Model Training: Retrain the prediction model using the prediction dataset. Call the prediction model's training interface and retrain the prediction model using the updated prediction dataset. The training method depends on the model used. For example, for a linear regression model, refit the regression coefficients; for a neural network model, perform several rounds of iterative updates to the weights. After training, save the updated model parameters, such as the weight file and coefficient matrix, for subsequent prediction. At the same time, record the training log, including training time, dataset size, model performance metrics (such as validation set error), etc.

[0222] Specifically, in this embodiment, incremental training is usually scheduled for late-night hours to avoid affecting real-time control during the day, such as 2:00-3:00 AM.

[0223] The system employs a delayed triggering mechanism: after detecting complete data at 22:00, training does not begin immediately, but instead a scheduled task is set to execute training at 2:00 AM. Training typically takes 5-10 minutes, consuming limited computing resources and not affecting other system functions. Furthermore, if data integrity verification fails (e.g., due to communication interruption or equipment failure resulting in missing critical data), the system executes a fault-tolerant process: skipping the current incremental training and using the current model parameters for subsequent predictions; simultaneously, the system logs the number of skips and the reasons. When the number of consecutive training skips reaches 3 days, the system automatically triggers an alarm mechanism, reminding maintenance personnel to check the data acquisition system (e.g., sensors, communication modules) via SMS, platform pop-ups, etc., ensuring the reliability of long-term data acquisition. This mechanism ensures system robustness while preventing poor-quality data from damaging model performance. Through the incremental training mechanism, the prediction model can always be based on data from the most recent 21 days, adapting promptly to changes in user electricity consumption patterns (e.g., seasonal changes, production mode adjustments), maintaining prediction accuracy. Actual observations show that without model updates, prediction errors gradually increase over time, potentially increasing by 20-30% after 3-6 months; however, with incremental training, prediction errors remain stable or even gradually decrease, demonstrating the effectiveness of continuous learning.

[0224] Specifically, in this embodiment, the present invention achieves closed-loop management of energy storage demand control throughout the entire process through a three-stage architecture. The multi-scenario configuration in the preparation stage ensures the system's adaptability to different date types; the rolling optimization in the execution stage enables real-time control and dynamic adjustment; and the incremental training in the learning stage ensures continuous optimization of the prediction model. Among these, prediction confidence assessment and adaptive safety margin provide robustness guarantees, and the robust optimization strategy... Complexity ensures real-time performance, while smoothing constraints and total quantity constraints balance equipment protection and resource utilization. Error history learning enables automatic parameter adjustment. Through the combination of these mechanisms, a complete, adaptive, and intelligent energy storage demand control method is formed, which is significantly superior to existing technologies in terms of robustness, real-time performance, and economy.

[0225] Specifically, in this embodiment, taking a pilot winter-time application scenario in a factory as an example, the factory's energy storage system capacity is only 240 kWh, while the peak load electricity consumption on ordinary working days is as high as 800-1200 kWh, with an energy storage coverage rate of only 20-30%, resulting in very limited regulation capabilities. More importantly, the demand electricity charges in this region adopt a special monthly settlement rule, that is, billing is based on the day with the highest demand in the month, rather than the monthly average demand. This means that even if a certain method performs well within 25 days, as long as there is a high demand failure on just one day, the electricity bill for the entire month will increase significantly. Therefore, the core indicator for evaluating demand control methods should be the monthly peak demand, rather than the monthly average demand. Based on this understanding, this invention conducted a 26-day (excluding rest days and holidays) experimental comparison of the average discharge strategy, the method of this invention, and the traditional MPC strategy in October 2025 at this factory. The specific indicators are shown in the table below.

[0226] Table 2 Experimental Comparison

[0227] Evaluation indicators Average discharge Method of the present invention Traditional MPC This invention is superior to average discharge. This invention is superior to MPC Peak monthly demand (kW) 256.0 243.8 267.3 -12.2 (-4.8%) -23.5 (-8.8%) Average monthly demand (kW) 227.5 222.0 223.4 -5.5 (-2.4%) -1.4 (-0.6%) Stability - Standard Deviation (kW) 28.0 26.1 31.4 -1.9 (-6.8%) -5.3 (-17.0%) 95th percentile demand (kW) 251.9 242.9 263.4 -9.0 (-3.6%) -20.5 (-7.8%) Monthly electricity cost (RMB / month) 12,800 12,190 13,365 -610 -1,175

[0228] As can be seen from the table, the method of this invention not only leads in terms of peak monthly demand, but also excels in monthly average demand, stability, and high-risk scenarios (95th percentile). Of particular note is the stability index, with a standard deviation of 26.1 kW, which is 6.8% lower than the average discharge strategy and 17.0% lower than the MPC strategy. The smaller the standard deviation, the more consistent the performance of the method is in different scenarios, and the lower the risk of extreme failure. The high standard deviation (31.4 kW) of the MPC strategy reflects its unstable characteristic: it can find extreme values ​​in some scenarios (e.g., only 229.4 kW on October 21st), but it fails severely in other scenarios (e.g., as high as 267.3 kW on October 10th). Analysis of 26 days of experimental data shows that the method of this invention significantly outperforms the average discharge strategy (reduced by 12.2 kW) and the traditional MPC strategy (reduced by 23.5 kW) in the key settlement indicator of monthly peak demand, resulting in annual electricity cost savings of up to 14,100 yuan (assuming a demand price of 50 yuan / kW·month). It also maintains a leading position in indicators such as monthly average demand and high-risk scenarios (95th percentile). From a technical perspective, the method of this invention, through the synergistic effect of four mechanisms—prediction confidence assessment, confidence threshold adaptation, total quantity constraint scaling, and load intensity tiered response—automatically switches to a conservative strategy in scenarios with large prediction errors, avoiding extreme failures of MPC. Therefore, the method of this invention does not "sacrifice performance for robustness," but rather achieves an organic unity of performance optimization, stability assurance, and engineering practicality in the most critical settlement metrics. Furthermore, the computation time of the method of this invention is only tens of milliseconds (seconds for MPC), requires no commercial solver (saving tens of thousands of yuan in licensing fees), and the algorithm is transparent and interpretable, making it suitable for embedded real-time deployment in practical industrial applications.

[0229] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present invention specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for adaptive energy storage demand control based on prediction confidence, characterized in that, Includes the following steps: S1. Preparation stage: Scene recognition and configuration selection are performed, the current date type is identified, and the corresponding configuration parameters are loaded according to the recognition results; S2, Execution Phase: A rolling closed-loop control mechanism is adopted, fully executing the "prediction-evaluation-adjustment-optimization-execution-feedback" closed-loop process in each first time period, achieving real-time and precise control of the energy storage system through dynamic iteration; the evaluation in the S2 closed-loop process specifically includes: S221. Construct a prediction error history sequence of length L, store prediction error data for L time steps, and calculate the prediction error; the prediction error is: in, For prediction error, For the prediction error history sequence The measured electricity consumption within a consecutive first time period, where the time step of each first time period corresponds to the end time. To predict electricity consumption; S222. Add the prediction error to the end of the prediction error history sequence, and determine whether the queue length exceeds L. If it does, delete the first element of the prediction error history sequence. S223. Based on the historical sequence of prediction errors, calculate... The average error at each time step and The average error at each time step respectively through average error and average error Conduct confidence level assessments and dynamically adjust safety margins; The adjustments in the closed-loop process of step S2 include: S231, through average error The confidence level assessment is conducted as follows: when When the value is ≤0.05, the confidence level is 1.0; When 0.05 < When the value is ≤0.10, the confidence level is 0.8; When 0.10 < When the value is ≤0.15, the confidence level is 0.6; when When the value is greater than 0.15, the confidence level is 0.3; S232, through average error The safety margin is dynamically adjusted, specifically as follows: when When the value is greater than 0.10, the prediction performance is considered to have deteriorated, and the safety margin is increased by 10%; specifically: ,in, For safety margin, This is the initial value for the safety margin. This represents the upper limit of the safety margin. when When the value is ≤0.05, the prediction performance is considered excellent, and the safety margin is reduced by 5%, specifically: , This is the lower limit of the safety margin; When 0.05 < When the value is ≤0.10, the prediction performance is considered stable and the safety margin remains unchanged. S233. The adjusted safety margin is applied to the load-average power correction, specifically as follows: in, For the first Predicted average power after time step adjustment For the first The original predicted average power at each time step; S3. Learning phase: After the execution phase is completed, the prediction model is incrementally updated using the actual operating data of the entire cycle on that day.

2. The energy storage demand control method based on predictive confidence adaptive control according to claim 1, characterized in that, The scene recognition in step S1 adopts a three-level priority determination mechanism, specifically as follows: Check the match between the current date and the preset holiday list; If a match is found, it is determined to be of the "Sunday / Holiday" type; If no match is found, the weekday code is checked. If the weekday code is 6, it is determined to be of type "Saturday"; otherwise, it is determined to be of type "regular weekday" by default.

3. The energy storage demand control method based on predictive confidence adaptive control according to any one of claims 1-2, characterized in that, The configuration parameters in step S1 are specifically as follows: Based on the identification results, the corresponding date type configuration parameters are automatically queried and loaded from the configuration parameter library. The configuration parameters include the decision start time, demand control period, control step size, and prediction reference historical data range.

4. The energy storage demand control method based on predictive confidence adaptive control according to any one of claims 1-2, characterized in that, The prediction in the closed-loop process of step S2 is specifically as follows: S211. Extract historical data from the database and input it into the prediction model; the prediction model adopts a linear regression model, neural network, or deep learning model for similar days; S212. The prediction model outputs the energy consumption sequence from the current time to the end time of demand control, and converts the energy consumption sequence into an average power sequence; the average power sequence is: in, for The average power consumption corresponding to the first time step of the time period. for The electricity consumption of the first time period step.

5. A method for adaptive energy storage demand control based on predictive confidence according to any one of claims 1-2, characterized in that, The optimization in the closed-loop process of step S2 is specifically as follows: S241. Construct an average discharge power baseline; specifically: Distribute the current remaining energy storage evenly to the remaining... At each time step, the average discharge power baseline is calculated, and the discharge plan is initialized with the average discharge power baseline to form a baseline scheme. The average discharge power baseline is: in, The average discharge power, This represents the current energy storage status. The remaining time steps. The time step interval; S242, Quantifying load intensity; specifically: Calculate the statistical characteristics of the load forecast sequence, including: The average load is: The standard deviation of the load is: in, For average load, The standard deviation of the load; Calculate the first based on the average load and load standard deviation. The load intensity at the time step; the load intensity is: in, For the first The load intensity at each time step; S243. Confidence-weighted load intensity grading fine-tuning; specifically: like If the value is ≤0.6, it indicates insufficient prediction reliability, and no fine-tuning will be performed. ,in, For the optimal discharge power sequence, As a baseline scheme, Confidence level; like A value greater than 0.6 indicates reliable prediction. For each time step, a graded response fine-tuning is performed based on load intensity, while a confidence-weighted control of the adjustment range is introduced; specifically: High load : The maximum discharge increase is 25%; Medium and high load : The maximum discharge increase is 18%; Mild high load : The maximum discharge increase is 10%; Normal load : Maintain the baseline; low load : The maximum discharge reduction is 25%. S244, Smoothing constraint processing; specifically: Set the maximum allowed change. A gradual adjustment strategy is adopted to achieve smooth constraints, specifically: the optimal discharge power sequence at the first time step. Keep it unchanged, every subsequent time step ,like Then let ,in, For the updated time step The optimal discharge power sequence; if Then let Otherwise, remain unchanged; S245, Discharge Total Constraint Calibration; specifically: The planned total discharge after smoothing constraints is calculated, and the total discharge is calibrated by uniformly scaling all time steps using a scaling factor; the planned total discharge is: The scaling factor is: The total discharge amount is: in, For the planned total discharge, Total discharge amount For time step interval, This represents the current energy storage status. This is the scaling factor; S246. Discharge power upper and lower limit constraints; the sequence of discharge power is as follows: in, It is a sequence of discharge power. For energy storage discharge power, [Minimum discharge power of energy storage, maximum discharge power of energy storage].

6. The energy storage demand control method based on predictive confidence adaptive control according to claim 5, characterized in that, The execution and feedback in the closed-loop process of step S2 are specifically as follows: S251, Discharge command executed, extract the first element of the discharge power sequence. As the discharge power command for the current time step, it is sent to the power conversion system of the energy storage system to control the energy storage devices according to... The discharge operation is performed, and the actual discharge capacity is calculated after the first time period; the actual discharge capacity is: in, This represents the actual discharge capacity. S252, Status Update and Feedback Learning; After the discharge operation is performed, the system updates the energy storage status. : The max function ensures the state is non-negative, and the result is compared with the detection value of the energy storage system. Compare them; if they are inconsistent, then let... At the same time, ensure ,in, The rated capacity of the energy storage system, These are the lower and upper limits of the preset state of charge for the energy storage system, respectively. If the system has already obtained the actual load data for the current first time period Then calculate the actual demand; the actual demand is: in, For actual demand, The first in the sequence of discharge power One element; Based on actual electricity consumption Calculate real-time error Update the historical prediction error series and recalculate the average error. and average error And based on the average error and average error Update the confidence level and safety margin to provide updated parameter support for the next time step optimization; S253. The system clock advances by one time step and determines whether the predicted end time has been reached. If not, return to step S2 to repeat the closed-loop process. If the time has been reached, execute step S3.

7. A method for adaptive energy storage demand control based on predictive confidence according to any one of claims 1-2, characterized in that, Before step S3, the method further includes: Set dual trigger conditions; specifically: Time condition verification: Determine whether the current time is greater than or equal to the end time of the prediction period; Data integrity verification: If the time condition is met, further determine whether there are actual load data records for the end of the forecast period and the entire cycle of the day; If the dual triggering conditions are met, incremental training is triggered; if the time condition is not met, the system remains in standby mode; if the data integrity condition is not met, the fault tolerance mechanism is activated.

8. A method for adaptive energy storage demand control based on predictive confidence according to any one of claims 1-2, characterized in that, The incremental training in step S3 employs a fixed-window-length rolling update strategy, specifically as follows: S31. Data Preprocessing and Standardization: Extract complete daily operational data from the database and perform data preprocessing and standardization procedures. Outlier detection, using 3 The criteria identify out-of-range abnormal data and determine the validity of the data; Missing values ​​are filled using linear interpolation for data with a missing rate of 5% or less; data with a missing rate greater than 5% are discarded. Feature normalization maps load data and associated features to interval; S32. Rolling update of the training dataset: Add the preprocessed data of the day to the prediction dataset and check the size of the prediction dataset. S33. Incremental training of the model: retrain the prediction model using the prediction dataset.

Images