Load prediction and intelligent real-time control system based on hybrid algorithm fusion

Abstract

The invention provides a load prediction and intelligent real-time control system based on hybrid algorithm fusion. Relates to the field of energy storage systems and energy scheduling, and comprises a data acquisition and processing module used for acquiring and preprocessing multi-source heterogeneous data of the energy storage system in multiple scenes to generate a structured time sequence feature set; the multi-source heterogeneous data comprises historical load data, time variables, environmental data, economic data and scene exclusive data; the power load prediction module outputs a load prediction result by integrating two different types of prediction models, and the two prediction models comprise a first prediction model based on historical sequence similarity and a second prediction model based on time decomposition; and the intelligent real-time control module is used for dynamically generating an energy storage system charging and discharging instruction and an energy scheduling strategy through a rolling optimization strategy based on a load prediction result and a preset target function, so that multi-target energy management is realized. According to the invention, the load response capability and prediction control precision of the energy storage system in multiple scenes are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy storage systems and energy scheduling, and particularly relates to a load prediction and intelligent real-time control system based on hybrid algorithm fusion. BACKGROUND

[0002] With the development of new power systems and the widespread deployment of renewable energy and energy storage systems, the load volatility and scheduling complexity in distributed scenarios have significantly increased. In typical application scenarios such as parks, construction sites, and construction waste resource processing, the load presents periodicity, strong peak fluctuation, and high uncertainty, which brings great challenges to load prediction and energy scheduling control.

[0003] Although existing technologies integrate different prediction models, there is a lack of collaborative optimization mechanism between the models, which cannot solve the inherent contradiction that a single model lacks the ability to capture sudden signals and long-term trends in non-stationary and multi-scenario coupled load sequences. In addition, the coupling between the prediction results and the control system is loose, and the control strategy relies on preset fixed thresholds, which leads to insufficient dynamic response capability of the energy storage system in real-time fluctuation scenarios with renewable energy access, resulting in control lag and low energy scheduling efficiency.

[0004] Therefore, it is urgent to build an intelligent system that can deeply integrate multiple traditional prediction algorithms with complementary principles and achieve close linkage between prediction results and control strategies. The system aims to solve the problems of insufficient single model capability, lack of model collaboration, and control response lag in existing methods in non-stationary load sequences, thereby improving the real-time scheduling efficiency and economy of the energy storage system in distributed energy scenarios. SUMMARY

[0005] To solve the above problems, the present application proposes a load prediction and intelligent real-time control system based on hybrid algorithm fusion, aiming to improve the load response capability and prediction control accuracy of the energy storage system in multiple scenarios.

[0006] To achieve the above purpose, the following technical solutions are implemented:

[0007] A load prediction and intelligent real-time control system based on hybrid algorithm fusion, comprising:

[0008] A data acquisition and processing module for acquiring and preprocessing multi-source heterogeneous data of the energy storage system in multiple scenarios to generate a structured time series feature set; the multi-source heterogeneous data includes historical load data, time variables, environmental data, economic data, and scenario-specific data;

[0009] The power load prediction module outputs a load prediction result by integrating two different types of prediction models, including a first prediction model based on historical sequence similarity and a second prediction model based on time decomposition.

[0010] The first prediction model is used to capture local mutations and non-stationary characteristics in the load sequence, adopts a weighted nearest neighbor algorithm with fast parameter tuning, realizes fast prediction through a distance ranking weighting mechanism and automatic parameter tuning, and obtains a first load prediction result in the future.

[0011] The second prediction model is used to extract long-term trends and periodic components in the load sequence, adopts a Holt-Winters exponential smoothing model, captures periodic characteristics by performing exponential smoothing and iterative updating on the three components of level, trend, and season, and fuses the updated level, trend, and season components to obtain a second load prediction result in the future.

[0012] The first load prediction result and the second load prediction result are linearly weighted and fused through a weighted fusion strategy to output a final load prediction result.

[0013] The intelligent real-time control module generates energy storage system charge and discharge instructions and energy scheduling strategies based on the load prediction result and real-time load state of the energy storage system, and realizes multi-objective energy management.

[0014] Further, the first prediction model based on historical sequence similarity is constructed using a weighted nearest neighbor algorithm with fast parameter tuning, specifically including:

[0015] The historical load sequence is non-overlappingly sampled through a sliding step n to generate historical sub-sequences with a window length of p;

[0016] The Euclidean distance between the current sub-sequence and the historical sub-sequences is calculated;

[0017] The first k neighbors are selected according to the Euclidean distance sorting, and weights are assigned based on the ranking, including: adopting a decay weighting rule based on the sorting position, so that the weight of the jth nearest neighbor decreases according to the harmonic series of 1 / j;

[0018] The first load prediction result is generated by weighted averaging according to the weights, and in this process, a distance consistency judgment mechanism is introduced: when multiple historical sub-sequences have the same distance value as the current sequence, a binary indicator function is set to retain their equivalent influence.

[0019] Further, the first prediction model based on historical sequence similarity realizes automatic parameter tuning through improved cross-validation, specifically including:

[0020] The training set and test set are split backwards according to the prediction step size, where the training set retains the historical continuous sequence and the test set corresponds to the future continuous prediction period.

[0021] Under the condition of satisfying the total data volume constraint, traverse the parameter combination, and constrain the sum of the historical window length and the number of neighbors to not exceed the available data capacity;

[0022] Calculate the average error of multi-round cross-validation for each parameter combination, select the parameter with the smallest average absolute percentage error as the optimal solution, and obtain the optimal historical window length and the optimal number of neighbors.

[0023] Furthermore, the second prediction model based on time decomposition is constructed using the Holt-Winters exponential smoothing model and employs a third-order component decoupling update mechanism, specifically including:

[0024] The horizontal component achieves dynamic benchmark calibration by weighted fusion of the current actual load observation value and historical trend. The horizontal component is the weighted average of the current observation value after removing the seasonal influence of the previous period and the horizontal extrapolation value of the previous period. The weight is controlled by the horizontal smoothing coefficient.

[0025] The trend component is estimated based on the difference between adjacent level values. A trend smoothing coefficient is used to control the fusion strength of historical trend continuity and recent change characteristics, so as to achieve gradual tracking of trend evolution.

[0026] The seasonal component constructs a periodic pattern by stripping the residual terms after removing the trend and level, and weights and fuses the current periodic phase observation with the historical in-phase factor, with the weight distribution of new and old information adjusted by the seasonal smoothing coefficient.

[0027] Furthermore, the fused and updated level, trend, and seasonal components are used to obtain the second load forecast result for the next r time steps. A multi-step forecast synthesis mechanism is employed, specifically including:

[0028] Extrapolate future base load based on current horizontal components and trend slope;

[0029] The seasonal factor for locating the target time point is determined by looping around the seasonal index;

[0030] The extrapolated base load value and seasonal factors are superimposed to generate the final second load forecast result.

[0031] Furthermore, the first prediction model based on historical sequence similarity includes:

[0032] The sequence library building unit is used to cut historical load data into subsequences with a window length of p;

[0033] The Euclidean distance calculation unit is used to calculate the Euclidean distance between the current sequence and the historical sequences;

[0034] The neighbor filtering and weighting unit is used to select the top k neighbors by sorting them by Euclidean distance, assign weights based on the ranking, and then weight and fuse the prediction results based on the weights.

[0035] The parameter optimization unit automatically selects the optimal window length and number of neighbors based on cross-validation.

[0036] Furthermore, the Holt-Winters exponential smoothing model includes three recursive update components: level, trend, and season, and outputs the results through a predictive synthesis module.

[0037] Furthermore, the intelligent real-time control module, based on the predicted load results and real-time load status of the energy storage system, generates charging and discharging commands and energy dispatching strategies for the energy storage system to achieve multi-objective energy management, including:

[0038] When the difference between the predicted load result and the current load is positive and greater than the first threshold, the charging mode is activated to start charging and the energy scheduling strategy corresponding to the charging mode is executed:

[0039] When the difference between the predicted load result and the current load is negative and less than the second threshold, the discharge power is calculated, the discharge mode is activated, and the load is discharged using the calculated discharge power.

[0040] When the difference between the predicted load result and the previous load is greater than or equal to the second threshold and less than or equal to the first threshold, the maintenance mode and the energy dispatch strategy corresponding to the maintenance mode are executed.

[0041] Furthermore, the energy dispatch strategies corresponding to the charging mode, discharging mode, and maintenance mode of each equipment module of the energy system include: in the charging mode, executing at least one action command in the following order of priority: switching to photovoltaic charging, enabling off-peak electricity price charging, and limiting charging power; in the discharging mode, executing at least one action command in the following order of priority: ensuring power supply to critical loads, discharging at peak electricity prices, and smoothing load fluctuations; in the maintenance mode, executing at least one action command in the following order of priority: photovoltaic direct-supply loads and battery standby.

[0042] Furthermore, the intelligent real-time control module includes:

[0043] The load pattern analysis unit is configured to extract features from the future load forecast sequence and match it with a preset load pattern strategy library to output a preliminary control strategy sequence.

[0044] The multi-model confidence fusion unit is configured to receive the historical sequence similarity measure calculated by the first prediction model and / or the trend and seasonal component strengths decomposed by the second prediction model, and to perform confidence assessment and adaptive adjustment on the preliminary control strategy sequence.

[0045] The dynamic cost decision unit is configured to fine-tune the adjusted control strategy based on the battery loss cost evaluation function, and finally generate charging and discharging commands and energy dispatch strategies with the goal of minimizing the overall operating cost.

[0046] Compared with the prior art, the present invention achieves the following beneficial effects:

[0047] 1. This invention proposes a weighted nearest neighbor prediction method with rapid parameter tuning: by constructing historical load time subsequences and calculating Euclidean distances, combined with a weighting mechanism based on distance ranking, a low-cost and high-efficiency preliminary load prediction is achieved. Simultaneously, a cross-validation optimization mechanism is introduced to automatically tune the historical window length and the number of neighbors, improving the overall model robustness and prediction performance.

[0048] 2. This invention introduces the Holt-Winters exponential smoothing model to model trend and seasonal variations: It decomposes the horizontal, trend and seasonal components of the time series in the form of an additive model, and models the three separately using the exponential smoothing method. This can effectively capture the periodic fluctuation patterns in typical scenarios and enhance the model's generalization ability in multi-period environments.

[0049] 3. This invention constructs a fusion prediction mechanism to improve overall prediction accuracy: the prediction results of the fast weighted nearest neighbor model and the Holt-Winters exponential smoothing model are weighted and linearly fused to give full play to the advantages of each model in short-term mutation capture and long-term trend fitting, so as to obtain a more stable and accurate load prediction output.

[0050] 4. This invention achieves deep coupling between prediction and control, enhancing the intelligence and robustness of system decision-making. Through a three-level processing mechanism of load pattern analysis, multi-model confidence fusion, and dynamic cost decision-making, the intermediate results and uncertainty measures of load forecasting are directly transformed into the basis for adjusting the control strategy. This overcomes the shortcomings of traditional threshold control, such as lag and rigidity, enabling the energy storage system to proactively and adaptively respond to complex and ever-changing grid conditions.

[0051] 5. This invention constructs a multi-objective optimization framework that balances economic efficiency and equipment lifespan. By introducing a concrete battery loss cost assessment function and incorporating it, along with electricity price costs and load smoothing benefits, into a comprehensive cost-based decision-making process, it can effectively extend the lifespan of energy storage batteries while ensuring economical operation, achieving a balance between short-term gains and long-term asset health.

[0052] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0053] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0054] Figure 1 This is a schematic diagram of a load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the system architecture of a load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to an embodiment of the present invention;

[0056] Figure 3 This is a flowchart illustrating the first prediction model according to an embodiment of the present invention;

[0057] Figure 4 This is a flowchart illustrating the second prediction model according to an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram of the control system architecture of a load forecasting and intelligent real-time control system based on hybrid algorithm fusion, according to an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, and not all embodiments. 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.

[0060] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0061] Figure 1 This is a schematic diagram of a load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to an embodiment of the present invention.Figure 2 This is a schematic diagram of the system architecture of a load forecasting and intelligent real-time control system based on hybrid algorithm fusion, according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, a load forecasting and intelligent real-time control system 100 based on hybrid algorithm fusion includes:

[0062] The data acquisition and processing module 110 is used to acquire multi-source heterogeneous data from multiple scenarios of the energy storage system and perform preprocessing to generate a structured time-series feature set; the multi-source heterogeneous data includes: historical load data, time variables, environmental data, economic data and scenario-specific data;

[0063] S1: Multi-source data acquisition and preprocessing

[0064] S1.1: Multi-source data acquisition

[0065] When applying energy storage systems to various scenarios (such as industrial parks, construction sites, and construction waste recycling centers), it is essential to first conduct an in-depth analysis of the power load characteristics of each application scenario. The load curves in different scenarios differ significantly: industrial parks typically exhibit periodic peak electricity consumption during office hours; construction sites experience strong load fluctuations due to the unpredictable start-up and shutdown of construction equipment; and waste recycling scenarios often experience peak loads due to phased changes in the processing flow.

[0066] Therefore, it is necessary to clarify the specific requirements of various scenarios for load forecasting accuracy and real-time control response speed in order to guide the design of subsequent models and control systems.

[0067] The data collection process should integrate multi-source heterogeneous data, including: historical load data: power consumption, electricity usage duration, daily / weekly / seasonal distribution; time variables: such as weekdays / non-working days, holidays, day and night; environmental data: temperature, humidity, rainfall, light intensity, etc., reflecting the impact of weather on energy consumption; economic data: electricity price fluctuation information, peak-valley electricity price structure; and scenario-specific data: such as construction waste disposal volume, transportation frequency, construction stage and schedule, etc., reflecting the impact of scenario behavior on load.

[0068] Finally, a multi-dimensional time-series feature dataset is constructed to drive load forecasting and control strategy modeling.

[0069] S1.2: Multi-source data preprocessing

[0070] Raw data often contains missing values, outliers, and inconsistent scales. The following processing steps are required first: Data cleaning: Remove duplicate, invalid, and erroneous records, interpolate or delete missing data; Normalization / standardization: Unify the value range of features with different scales to improve the convergence speed and stability of the model.

[0071] Feature engineering: Extracting time features (hourly, daily, weekly, and holiday labels); constructing weather features (such as combining humidity and temperature to form a cooling and heating load index); encoding equipment status (such as the activation status of construction equipment and the operation stage of processing lines); using methods such as principal component analysis, mutual information, and correlation coefficients to select features, retaining the variables most strongly correlated with the load, and using the features after this feature selection process to supply the two prediction models.

[0072] Finally, model input data with a uniform format and controllable quality is generated for subsequent prediction model construction and training.

[0073] The power load forecasting module 120 outputs load forecasting results by integrating two different types of forecasting models, namely: a first forecasting model based on historical sequence similarity and a second forecasting model based on time decomposition.

[0074] The first prediction model captures local abrupt changes and non-stationary features in the load sequence. It employs a weighted nearest neighbor algorithm with fast parameter tuning, using a distance ranking weighting mechanism and automatic parameter tuning to achieve rapid prediction and obtain the first load prediction result. The second prediction model extracts the long-term trend and periodic components in the load sequence. It uses the Holt-Winters exponential smoothing model, performing exponential smoothing and iterative updates on the horizontal, trend, and seasonal components respectively to capture periodic features. The updated horizontal, trend, and seasonal components are then fused to obtain the second load prediction result. The weighted fusion unit performs linear weighted fusion of the first and second load prediction results using a weighted fusion strategy to output the final load prediction result. Furthermore, the power load prediction module 120 includes:

[0075] S2: Load Forecasting Model Construction

[0076] The power load forecasting module 120 comprises two sub-models and one fusion unit. The construction process specifically includes the following steps:

[0077] S2.1: Construction of the first prediction model based on historical sequence similarity

[0078] To improve the initial prediction accuracy of the system under different load scenarios, this invention proposes a weighted nearest neighbor (KNN) method for fast parameter tuning. The algorithm is briefly described as follows: 1. Slide sampling in the historical sequence with a step size n to construct a candidate subsequence set; 2. Calculate the similarity between the current subsequence and the candidate set based on Euclidean distance, and select the optimal k neighbors; 3. Calculate the load prediction value for the next step n according to the weighting rule, as the initial prediction reference input for the system. This method provides efficient predictive prior information for subsequent control decisions through low-cost approximate inference, significantly improving the overall system prediction accuracy and response efficiency.

[0079] The first prediction model based on historical sequence similarity employs a weighted nearest neighbor algorithm with fast parameter tuning. It achieves rapid prediction through a distance ranking weighting mechanism and automatic parameter tuning, yielding the first predicted load result. It includes: a sequence library construction unit for segmenting historical load data into subsequences with a window length of p; a Euclidean distance calculation unit for calculating the Euclidean distance between the current sequence and historical sequences; a neighbor selection and weighting unit for selecting the top k neighbors based on Euclidean distance, assigning weights based on ranking, and fusing the prediction results based on these weights; and a parameter optimization unit for automatically selecting the optimal window length and number of neighbors based on cross-validation.

[0080] Figure 3 This is a flowchart illustrating the first prediction model according to an embodiment of the present invention. Figure 3 As shown, the detailed algorithm flow is as follows:

[0081] S2.1.1 Feature Construction and Prediction Mechanism

[0082] By performing non-overlapping sampling on the historical load sequence using a sliding step size n, historical subsequences with a window length of p are generated. Specifically, the multidimensional load time series data is divided into historical subsequences of length p in chronological order:

[0083]

[0084] in, : The i-th historical subsequence, representing a load data segment that traces back p time steps from time point i; p: the length of the historical window, i.e., the number of time steps contained in each input sequence, used to characterize the historical load range of the current forecast reference.

[0085] This sequence serves as the input feature for the current time step; the corresponding prediction target is the value of its subsequent n time steps:

[0086]

[0087] The historical true load sequence for the i-th time point in time n steps ahead is used to train KNN, serving as the target prediction sequence and also as a historical subsequence. The corresponding subsequent actual load value sequence; n: prediction step size, indicating the load situation at n future time points each time prediction.

[0088] The prediction result for the current time T is the same as in the history. The weighted average output value of the k most similar subsequences. In the weighted averaging process, a decaying weighting rule based on ranking position is used, such that the weight of the j-th nearest neighbor decreases by a harmonic series of 1 / j. Simultaneously, a normalization factor is set to ensure that the total weight is inversely proportional to the number of neighbors k. In the distance-ranked (harmonic series decay) weighted averaging process, a distance consistency indicator function is introduced. This is used to handle situations where multiple historical subsequences have the same distance value as the current sequence, ensuring their equivalent influence.

[0089] The method employs time-series cross-validation, dividing the data into 5 folds with a prediction step size of n=4. Each fold has a training set length of ≥72h. The parameter combination iterates through p∈[24,48,72] and k∈[5,10,15]. The MAPE threshold is set to 5% to terminate the search.

[0090] The specific expression for generating the first load forecast value using the weighted mechanism is as follows:

[0091]

[0092] : The first load prediction result for the current time point T over the next n time steps (KNN output); k: the number of nearest neighbors, i.e., the number of historical subsequences used for reference; : Neighbor distance consistency indicator function, which is 1 when the distance between neighbor i and neighbor j is equal, and 0 otherwise; : The true target prediction value corresponding to the i-th similar subsequence (i.e., the load sequence for the next n time steps); : with the current subsequence Similar historical subsequence indexes; The weighting coefficient of the j-th neighbor reflects the degree of contribution of that neighbor to the weighted average, and is defined as:

[0093]

[0094] : Weighted values ​​based on distance sorting, the weight of the j-th nearest neighbor, the closer the distance (the higher the ranking), the greater the weight; Apply less weight to lower-ranked neighbors to achieve a distance-sensitive weighted average; : Normalization factor, to ensure that the total weight is within a controllable range; The indicator function representing consistent neighbor distance ranking is used to determine whether neighbor i and neighbor j have the same distance value. This function handles the case of duplicate distances among multiple neighbors, preserving their influence in the weighted average, and is defined as:

[0095]

[0096] : Indicates the current subsequence with historical subsequence The distance between them is usually expressed in Euclidean form; This is a binary indicator function that returns 1 if the condition is true, and 0 otherwise. More specifically, it is calculated as follows: Otherwise, 0, where ε is a preset small tolerance threshold (e.g., 1e-3) to ensure that equivalent weights are retained when the distances are the same.

[0097] S2.1.2: Subsequence Similarity Calculation and Model Parameter Tuning

[0098] Specifically, the distance between subsequences is measured using Euclidean distance, and the calculation formula is as follows:

[0099]

[0100] : subsequence and The similarity distance between them; : These are the i-th and j-th elements in the historical load sequence, respectively, with lengths of . subsequence; Euclidean distance norm, used to measure the geometric distance between two vectors.

[0101] This distance is used to select the historical reference subsequence that is most similar to the current sequence to be predicted.

[0102] By calculating the Euclidean distance between the current subsequence and the historical subsequences, and introducing a distance consistency judgment mechanism: when multiple historical subsequences have the same distance value as the current sequence, their equivalent influence is preserved by setting a binary indicator function; the top k neighbors are selected according to Euclidean distance, and weights are assigned based on the ranking.

[0103] S2.1.3: Model Training

[0104] To improve model robustness, the first prediction model based on historical sequence similarity introduces an automatic parameter selection mechanism based on an improved cross-validation method to achieve automatic parameter tuning. Specifically, this includes: rolling the training set and test set backward according to the prediction step size, where the training set retains continuous historical sequences and the test set corresponds to continuous future prediction periods; traversing parameter combinations under the condition of satisfying the total data volume constraint, and constraining the sum of the historical window length and the number of neighbors to not exceed the available data capacity; calculating the average error of multi-round cross-validation for each parameter combination, and selecting the parameter with the smallest average absolute percentage error as the optimal solution to obtain the optimal historical window length and the optimal number of neighbors.

[0105] Specifically, the time series data is divided into a training set and a corresponding test set. The training set is as follows:

[0106]

[0107] : The training set in the i-th cross-validation; Time point The load observations; T: the total length of the complete time series; n: the prediction step size, i.e., the time span of each prediction; i: the round number index of the current cross-validation; this training set is used for model learning in the i-th iteration.

[0108] The test set is:

[0109]

[0110] : The test set corresponding to the i-th cross-validation; : indicates the training set Load data for the next n time points;

[0111] This test set is used to evaluate the model's predictive ability on unknown data.

[0112] S2.1.4 Loss Function Design

[0113] The training objective is to minimize the mean absolute percentage error (MAPE).

[0114]

[0115] The final selected optimal parameter combination, where: Optimal historical window length; : Optimal number of neighbors; argmin: Find the parameter pairs that minimize the objective function; In parameter combinations The mean absolute percentage error obtained below; I: the number of folds in cross-validation, i.e., the number of times the training set / test set is split; : No. Compromise, in parameters The prediction error is as follows;

[0116] This process is used to automatically select the historical window length and the number of neighbors that minimize the error, thereby improving the model's adaptability and generalization ability.

[0117] To ensure the training set There are enough data points to construct a subsequence of length p. Set the following constraints: The length of the training sequence is determined by p, k, and k is limited only by the number of available neighbors and cannot exceed the number of candidate neighbors that the training set can generate. During cross-validation, p and k iterate through positive integers while satisfying the minimum length requirement of the available training set.

[0118] This constraint ensures that, provided sufficient training and testing samples are allocated, the values ​​of p and k do not exceed the range of available data, thus guaranteeing the effectiveness of model construction and evaluation.

[0119] S2.2 Construction of the Second Prediction Model Based on Time Decomposition

[0120] To accurately capture the long-term trend and periodic variation characteristics in the load series and effectively complement them with the short-term local fluctuation characteristics, this invention introduces the Holt-Winters exponential smoothing model, which is specifically used to extract and model the trend component and seasonal component in the load series.

[0121] The second prediction model based on time decomposition adopts the Holt-Winters exponential smoothing model. It captures periodic characteristics by exponentially smoothing and iteratively updating the horizontal, trend, and seasonal components respectively. The updated horizontal, trend, and seasonal components are then fused to obtain the future second load prediction results.

[0122] Furthermore, the second prediction model based on time decomposition is constructed using the Holt-Winters exponential smoothing model, which employs a third-order component decoupling update mechanism. Specifically, the Holt-Winters model includes three recursive update components: horizontal, trend, and seasonal. It adopts an additive model form, exponentially smoothing the horizontal, trend, and seasonal components respectively, and then outputs the results through the prediction synthesis module.

[0123] Figure 4 This is a flowchart illustrating the second prediction model according to an embodiment of the present invention, as shown below. Figure 4 As shown, the specific construction process of the second prediction model includes:

[0124] 1. The horizontal component achieves dynamic benchmark calibration by weighted fusion of the current actual load observations and historical trends. Specifically, the horizontal component is the current observation value after removing the seasonal effects of the previous period. Compared with the previous period's horizontal extrapolation value ( The weighted average of ) is controlled by the horizontal smoothing coefficient ρ.

[0125] Specifically, horizontal component updates:

[0126]

[0127] The smoothed level value at the current time t represents the base load level; The actual observed load value at the current time t; Previous period (time) The corresponding seasonal factors; The smoothing level value at the previous time step; : The trend component of the previous moment; ρ: Horizontal smoothing coefficient, with a value range of [0,1], which determines the weighting of the current observation and the historical trend on the horizontal estimate.

[0128] 2. The trend component is estimated based on the difference between adjacent level values. The trend smoothing coefficient is used to control the fusion strength of historical trend continuity and recent change characteristics, so as to realize the gradual tracking of trend evolution.

[0129] Specifically, trend component updates:

[0130]

[0131] The trend component at the current time t reflects the changing trend of the load level; The difference between the current level and the previous level is used to estimate the slope of the trend. Trend smoothing coefficient: controls the response speed of trend updates, with a value range of [0,1].

[0132] 3. The seasonal component constructs a periodic pattern by stripping the residual terms after removing the trend and level, and weights and fuses the current periodic phase observation with the historical in-phase factor, with the weight distribution of new and old information adjusted by the seasonal smoothing coefficient.

[0133] Specifically, seasonal portion updates:

[0134]

[0135] The seasonal component at the current time t reflects the impact of periodic fluctuations on the load; The remaining portion after removing the trend and smoothing levels is used as a seasonal estimate; : Seasonal component at the same time point in the previous cycle; γ: Seasonal smoothing coefficient, which controls the influence of new observations on seasonal factors, with a value range of [0,1].

[0136] Among them, ρ, τ, and γ are optimized using the first 70% of historical training data, with initial values ​​set to ρ=0.2. =0.1, γ=0.1, and the L-BFGS-B algorithm is used to iterate until convergence.

[0137] 4. A multi-step forecasting synthesis mechanism is adopted to fuse the updated horizontal, trend, and seasonal components to obtain the second load forecast result for the next r time steps, including: extrapolating the future base load based on the current horizontal component and trend slope; locating the seasonal factor at the target time point by looping back through the seasonal index; and superimposing the extrapolated base load value and the seasonal factor to generate the final second load forecast result.

[0138] Specifically, the second load forecast for the next r steps:

[0139]

[0140] The second load forecast value for time point t+r predicted by the Holt-Winters model; : A linear growth estimate for the next r steps based on the current trend; : Seasonal component corresponding to the forecast time, the periodic adjustment effect on future load; r: forecast step size, representing the number of time steps to forecast to the future; m: seasonal cycle length, usually set according to load cycle characteristics (e.g., 24 for daily cycle, 168 for weekly cycle), can be defined according to the scenario, for example: in a park scenario, m=24 (daily cycle), in a construction site due to work shifts, m=48 (half-day cycle), in a waste disposal scenario, m=168 (weekly cycle); V=Floor (r - 1) / m+ 1, wrapped around the seasonal index (t + r - m) V) The latest seasonal factor estimate corresponding to the target location time point t+r .

[0141] After model initialization, the above parameters are iteratively optimized using historical load data. S2.3 Prediction results are then fused and output.

[0142] To achieve synergistic improvement in forecast accuracy among different methods, a multi-model weighted fusion strategy can be introduced. This strategy involves linearly weighting and fusing the first and second load forecast results to output the final load forecast result. :

[0143]

[0144]

[0145] The final prediction result after fusion is used for load forecasting at time t+r; The weight coefficients of the three models satisfy the following conditions: , used to control the contribution ratio of each model's prediction results to the final output; r: prediction step size, indicating that the predicted load value is at the r-th time point in the future; The first load forecast result of step S2.1 (for) Take the corresponding step size value). This is the second load forecast result from step S2.2. This value emphasizes the identification of medium- to long-term load trends and is applicable to periodic scenarios. The load value predicted at time T by the weighted nearest neighbor method at step r is also... The r-th element in the sequence, i.e., the KNN prediction value at time T+r, comes from a nonparametric method based on the similarity of historical subsequences, which is good at fitting short-term local patterns.

[0146] In this embodiment of the invention, the weighting coefficients in the weighted fusion strategy ( , The parameter is a dynamic adaptive parameter that is adjusted in real time based on the prediction time scale and historical prediction errors. The adjustment rule is as follows: = f(short-term forecast error), =g(long-term prediction error), thereby achieving adaptive complementarity of model advantages.

[0147] The specific calculation method for the dynamic adaptive weights is as follows:

[0148] For short-term forecasts (such as the next 1-4 hours), let = 0.7, = 0.3.

[0149] For medium- to long-term forecasts (such as the next 5-24 hours), let = 0.3, = 0.7. Simultaneously, fine-tuning is performed based on the mean absolute error (MAE) of each model's predictions from the previous period:

[0150]

[0151]

[0152] in, and These are the mean absolute errors of the first and second prediction models in the previous period, respectively. Finally, for... and Perform normalization to ensure that the sum is 1.

[0153] The intelligent real-time control module 130 generates charging and discharging commands and energy dispatching strategies for the energy storage system based on the predicted load results and real-time load status of the energy storage system, thereby achieving multi-objective energy management.

[0154] Optionally, a three-stage decision-making rule can be adopted to achieve energy dispatch and multi-objective energy management:

[0155] When the difference between the predicted load result and the current load is positive and greater than a first threshold (optional threshold) When the power is 10kW, the charging power is calculated, the charging mode is turned on and the calculated charging power is used, and the energy dispatch strategy corresponding to the charging mode is executed.

[0156] Calculate charging power :

[0157]

[0158] in, The preferred adjustment coefficient is... ; That is, a time period of 15 minutes; The maximum charging power is set to, for example, 200kW; The safe charging limit is set, for example, 90%; SOC is the ratio of current capacity to rated capacity; This is the current load value. This is the load forecast value; The nominal energy capacity represents the maximum total energy that a battery can store under standard conditions. For example, the nominal energy capacity of a battery is 500 kWh.

[0159] When the difference between the predicted load result and the previous load is negative and less than the second threshold (optional threshold) When the power is -10kW, calculate the discharge power, start the discharge mode and discharge with the calculated discharge power, and execute the energy dispatch strategy corresponding to the discharge mode.

[0160] Calculate discharge power :

[0161]

[0162] in, For example, the discharge efficiency is set to 0.92; The lower limit of safe discharge is, for example, 20%; SOC is the ratio of current capacity to rated capacity. This is the discharge coefficient, for example, taken as 0.8; This is the current load value. This represents the load forecast.

[0163] Preferably, the aforementioned threshold ΔL is based on the standard deviation of the scenario load. Adaptive adjustment: ,in The standard deviation of historical load is used, with a coefficient of 0.15, verified based on scenario experiments. Constraints are set for charging and discharging power calculation: Maximum charging power: (Retain a 20% power margin to cope with sudden loads); Discharge depth protection: Full discharge is allowed when SOC ≥ 30%.

[0164] When the difference between the predicted load result and the previous load is greater than or equal to the second threshold and less than or equal to the first threshold, the maintenance mode is executed, and the energy dispatch strategy corresponding to the maintenance mode is executed.

[0165] For example, when a certain energy storage system is at 10:00, the measured current load is 180 kW; the predicted load is 220 kW (+22%), the current SOC is 65%, and the measured photovoltaic output is 50 kW.

[0166] Based on the above parameters, the charging power can be calculated as follows: . Since the photovoltaic output is 50kW > the charging demand (i.e., charging power) If the power is 48kW, the system will switch to the instruction to charge the photovoltaic system with a charging power of 48kW.

[0167] Optionally, in embodiments of the present invention, the intelligent real-time control module 130 can make forward-looking decisions by analyzing the overall load forecast pattern over a future period and formulating the optimal control strategy sequence in advance. Specifically, it includes:

[0168] The load pattern analysis unit 131 is configured to extract features from the future load forecast sequence and match it with a preset load pattern strategy library to output a preliminary control strategy sequence.

[0169] The feature extraction performed by the load pattern analysis unit 131 includes at least: the overall change slope of the future load sequence, the fluctuation intensity, and whether it includes peak and valley periods associated with electricity price signals.

[0170] S3.1: Forward-looking analysis of load patterns

[0171] The controller receives a load forecast sequence for the next N time steps from the power load forecasting module. .

[0172] Feature extraction: The system calculates multiple features of the predicted sequence in real time, including: Slope: the trend component based on the HW model. Calculate the average slope of change for future periods; Fluctuation: Calculate the standard deviation or average volatility of the series, representing uncertainty; PeakValley: Determine whether future periods contain obvious peak and trough time points (in conjunction with electricity price signals).

[0173] Morphological matching: The extracted feature vectors are matched against a pre-defined policy library. The policy library contains various typical load patterns and their corresponding optimal control policy sequences. For example:

[0174] Type A ("Ready to Go"): Characterized by a current low load and a significantly positive trend slope, indicating an imminent peak electricity price period. Matching Strategy: Immediately initiate maximum power charging until the SOC reaches 95% of the safe limit.

[0175] Type B ("High-Level Oscillation"): Characterized by persistently high load with strong volatility. Matching Strategy: Employs a smoothing filter mode, where discharge power follows load fluctuations inversely to smooth the net load curve, rather than maximizing returns.

[0176] Type C ("Continuous Decline"): Characterized by a continuous decrease in load and a drop in electricity prices. Matching strategy: Enter standby or extremely low-power "float charging" state to conserve battery energy in preparation for subsequent sudden changes.

[0177] The load pattern strategy library is constructed as follows: Historical load data and corresponding optimal control strategies for at least one year are collected. The load curve patterns are then divided into five typical categories using the K-means clustering algorithm. A basic control strategy template is derived for each category through manual labeling or reinforcement learning optimization, thus forming the strategy library. In addition to the aforementioned patterns, the strategy library can also include "nighttime stable type" (corresponding to maintenance mode) and "peak-island type" (corresponding to ensuring power supply to critical loads), etc.

[0178] The multi-model confidence fusion unit 132 is configured to receive the historical sequence similarity measure calculated by the first prediction model and / or the trend and seasonal component strengths decomposed by the second prediction model, and to perform confidence evaluation and adaptive adjustment on the preliminary control strategy sequence.

[0179] The working logic of the multi-model confidence fusion unit 132 is as follows: when the similarity metric of historical sequences is lower than the preset threshold, the upper limit of the charge and discharge power instruction is automatically reduced; when the intensity of the trend component exceeds the preset threshold, the power response margin is reserved in advance.

[0180] S3.2 Multi-model intermediate result fusion decision

[0181] The decision-making mechanism of this invention deeply integrates the intermediate calculation results of two prediction models to achieve more accurate and robust judgments:

[0182] Confidence from the first prediction model (KNN): When the KNN model searches for similar sequences, the average Euclidean distance between its k nearest neighbors and the current sequence. It is used as a confidence index for decisions. If An excessively large value indicates low historical similarity and high uncertainty in the prediction results. In this case, the controller will tend to adopt a more conservative control strategy (such as reducing the charging and discharging power limit). For example, when the historical sequence similarity metric (i.e., mean Euclidean distance) is too large... When the distance is higher than the 85th percentile of the historical distance distribution, the upper limit of the charge and discharge power will be automatically reduced to 50% of the original value.

[0183] Component analysis from the second prediction model (HW): Strength of trend components decomposed using the HW model and seasonal components For example, when When the value is very high, it indicates that the load is in a period of rapid change, and the controller will reserve more power margin in advance to cope with sudden changes; while when When the seasonal factor is displayed as a typical "midday peak," the controller will combine time information to complete charging preparations in advance before noon. For example, when the trend component strength... If the trend strength exceeds twice the standard deviation of the baseline trend for three consecutive cycles, reserve an additional 20% power response margin in advance.

[0184] The dynamic cost decision unit 133 is configured to fine-tune the adjusted control strategy based on the battery loss cost evaluation function, and finally generate charging and discharging instructions and energy scheduling strategies with the goal of minimizing the overall operating cost.

[0185] S3.3 Policy Fine-Tuning Based on Dynamic Cost Function

[0186] The system has a built-in battery damage cost assessment function, Cost(P, SOC). This function is a simplified model based on electrochemical experience, used to estimate the cumulative damage to the battery under different charge and discharge powers P at different SOCs in real time.

[0187] Specifically, the battery degradation cost assessment function Cost(P, SOC) takes the following form:

[0188]

[0189] Where P is the instantaneous charging and discharging power (discharging is positive, charging is negative). Here, SOC is the rated power, SOC is the current state of charge, SOC_min is the minimum allowable SOC, and ΔT is the control period. (Coefficient) , , , These are parameters calibrated through aging tests based on the battery model. For example, for lithium iron phosphate batteries, their typical value range is: =1e-5, =0.5, =2e-3, =0.1.

[0190] Before issuing the final control command, the system calculates the overall cost of the candidate strategies:

[0191]

[0192] The calculation methods for each cost and benefit item are as follows:

[0193] Electricity cost: ,in, Let be the grid electricity price at time t+i. Let t+i be the power purchased from the grid (purchased is positive, sold is negative).

[0194] Load smoothing benefits: ,in, The net load at time t+i is (total load - renewable energy output - energy storage output). K represents the average net load over the forecast period, and K is the revenue coefficient (which can be set according to policy or demand, for example, 100 yuan / (kW·h)). This revenue term is inversely proportional to the smoothness of the net load; the smoother the net load, the smaller the penalty value of this term, and the lower the total cost.

[0195] α, β, and γ are weighting coefficients used to adjust the relative importance of different objectives. Typically, α = 1.0, β = 0.5, and γ = 0.1.

[0196] By comparing TotalCost under different fine-tuning strategies, an optimal command that balances economy and battery life can be selected. For example, even if the form factor indicates full discharge, if the current SOC is low, the Cost will be high, and the system may automatically limit the discharge power to a range more friendly to the battery.

[0197] In some embodiments, such as Figure 5This is a schematic diagram of the control system architecture of a load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to an embodiment of the present invention. The control system includes: an energy storage system: battery SOC, power limiting; a converter module: AC / DC conversion control; an energy switching module: switching the power supply path of the external power grid, photovoltaic, and energy storage; a local controller and a higher-level energy management system: receiving instructions, monitoring status, summarizing data, etc. Further, the energy dispatching strategies corresponding to the charging mode, discharging mode and maintenance mode of each device module include: in the charging mode, executing at least one action instruction in the following order of priority: switching to photovoltaic charging, enabling off-peak electricity price charging and limiting charging power; in the discharging mode, executing at least one action instruction in the following order of priority: ensuring power supply to critical loads, peak electricity price discharging and smoothing load fluctuations; in the maintenance mode, executing at least one action instruction in the following order of priority: photovoltaic direct supply to load and battery standby. Specifically, in the maintenance mode, the system executes the following energy dispatching strategies: (1) giving priority to using photovoltaic output to directly supply the load. (2) If there is surplus photovoltaic power and the current SOC is lower than the preset floating charging limit (e.g., 85%), the battery will be charged with the minimum floating power (e.g., 0.02C); if the SOC has reached the limit, the photovoltaic power will be abandoned. (3) If the photovoltaic power is insufficient, the grid will supplement the deficit and the battery will remain in standby mode without charging or discharging.

[0198] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0199] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0200] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A load forecasting and intelligent real-time control system based on hybrid algorithm fusion, characterized in that, include: The data acquisition and processing module is used to acquire multi-source heterogeneous data from various scenarios of the energy storage system and perform preprocessing to generate a structured time-series feature set. The multi-source heterogeneous data includes: historical load data, time variables, environmental data, economic data, and scenario-specific data; The power load forecasting module outputs load forecasting results by integrating two different types of forecasting models, namely: a first forecasting model based on historical sequence similarity and a second forecasting model based on time decomposition. The first prediction model is used to capture local mutations and non-stationary features in the load sequence. It adopts a weighted nearest neighbor algorithm with fast parameter tuning, and achieves fast prediction through distance ranking weighting mechanism and automatic parameter tuning to obtain the first load prediction result in the future. The second prediction model is used to extract the long-term trend and periodic components in the load sequence. It adopts the Holt-Winters exponential smoothing model, and captures the periodic characteristics by exponentially smoothing and iteratively updating the horizontal, trend and seasonal components respectively. The updated horizontal, trend and seasonal components are then fused to obtain the future second load prediction result. The first load forecast result and the second load forecast result are linearly weighted and fused using a weighted fusion strategy to output the final load forecast result. The intelligent real-time control module generates charging and discharging commands and energy dispatching strategies for the energy storage system based on the load forecast results and real-time load status of the energy storage system, thereby achieving multi-objective energy management.

2. The load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to claim 1, characterized in that, in, The first prediction model based on historical sequence similarity is constructed using a weighted nearest neighbor algorithm with fast parameter tuning, specifically including: By using a sliding step size n, non-overlapping sampling is performed on the historical load sequence to generate a historical subsequence with a window length of p; Calculate the Euclidean distance between the current subsequence and the historical subsequences; The k nearest neighbors are selected according to the Euclidean distance and weights are assigned based on the ranking, including: using a decay weighting rule based on the sorting position, so that the weight of the j-th nearest neighbor decreases by a harmonic series of 1 / j. The first load prediction result is generated by weighted average of the weights. During this process, a distance consistency judgment mechanism is introduced: when multiple historical subsequences have the same distance value as the current sequence, their equivalent influence is retained by setting a binary indicator function.

3. The load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to claim 2, characterized in that, The first prediction model based on historical sequence similarity achieves automatic parameter tuning through improved cross-validation, specifically including: The training set and test set are split backwards according to the prediction step size, where the training set retains the historical continuous sequence and the test set corresponds to the future continuous prediction period. Under the condition of satisfying the total data volume constraint, traverse the parameter combination, and constrain the sum of the historical window length and the number of neighbors to not exceed the available data capacity; Calculate the average error of multi-round cross-validation for each parameter combination, select the parameter with the smallest average absolute percentage error as the optimal solution, and obtain the optimal historical window length and the optimal number of neighbors.

4. The load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to claim 1, characterized in that, The second prediction model based on time decomposition is constructed using the Holt-Winters exponential smoothing model and employs a third-order component decoupling update mechanism, specifically including: The horizontal component achieves dynamic benchmark calibration by weighted fusion of the current actual load observation value and historical trend. The horizontal component is the weighted average of the current observation value after removing the seasonal influence of the previous period and the horizontal extrapolation value of the previous period. The weight is controlled by the horizontal smoothing coefficient. The trend component is estimated based on the difference between adjacent level values. A trend smoothing coefficient is used to control the fusion strength of historical trend continuity and recent change characteristics, so as to achieve gradual tracking of trend evolution. The seasonal component constructs a periodic pattern by stripping the residual terms after removing the trend and level, and weights and fuses the current periodic phase observation with the historical in-phase factor, with the weight distribution of new and old information adjusted by the seasonal smoothing coefficient.

5. The load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to claim 4, characterized in that, The merged and updated horizontal, trend, and seasonal components yield a second load forecast result for the next r time steps. A multi-step forecast synthesis mechanism is employed, specifically including: Extrapolate future base load based on current horizontal components and trend slope; The seasonal factor for locating the target time point is determined by looping around the seasonal index; The extrapolated base load value and seasonal factors are superimposed to generate the final second load forecast result.

6. The load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to claim 3, characterized in that, in, The first prediction model based on historical sequence similarity includes: The sequence library building unit is used to cut historical load data into subsequences with a window length of p; The Euclidean distance calculation unit is used to calculate the Euclidean distance between the current sequence and the historical sequences; The neighbor filtering and weighting unit is used to select the top k neighbors by sorting them by Euclidean distance, assign weights based on the ranking, and then weight and fuse the prediction results based on the weights. The parameter optimization unit automatically selects the optimal window length and number of neighbors based on cross-validation.

7. The load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to claim 5, characterized in that, in, The Holt-Winters exponential smoothing model includes three recursive update components: level, trend, and seasonality, and outputs results through a predictive synthesis module.

8. The load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to claim 7, characterized in that, in, The intelligent real-time control module, based on the predicted load results and real-time load status of the energy storage system, generates charging and discharging commands and energy dispatching strategies for the energy storage system, realizing multi-objective energy management, including: When the difference between the predicted load result and the current load is positive and greater than the first threshold, the charging mode is activated to start charging and the energy scheduling strategy corresponding to the charging mode is executed: When the difference between the predicted load result and the current load is negative and less than the second threshold, the discharge power is calculated, the discharge mode is activated, and the load is discharged using the calculated discharge power. When the difference between the predicted load result and the previous load is greater than or equal to the second threshold and less than or equal to the first threshold, the maintenance mode and the energy dispatch strategy corresponding to the maintenance mode are executed.

9. The load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to claim 8, characterized in that, The energy dispatch strategies corresponding to the charging mode, discharging mode, and maintenance mode of each equipment module of the energy system include: in the charging mode, executing at least one action command in the following order of priority: switching to photovoltaic charging, enabling off-peak electricity price charging, and limiting charging power; in the discharging mode, executing at least one action command in the following order of priority: ensuring power supply to critical loads, discharging at peak electricity prices, and smoothing load fluctuations; in the maintenance mode, executing at least one action command in the following order of priority: photovoltaic direct-supply loads and battery standby.

10. The load forecasting and intelligent real-time control system based on hybrid algorithm fusion according to claim 9, characterized in that, The intelligent real-time control module includes: The load pattern analysis unit is configured to extract features from the future load forecast sequence and match it with a preset load pattern strategy library to output a preliminary control strategy sequence. The multi-model confidence fusion unit is configured to receive the historical sequence similarity measure calculated by the first prediction model and / or the trend and seasonal component strengths decomposed by the second prediction model, and to perform confidence assessment and adaptive adjustment on the preliminary control strategy sequence. The dynamic cost decision unit is configured to fine-tune the adjusted control strategy based on the battery loss cost evaluation function, and finally generate charging and discharging commands and energy dispatch strategies with the goal of minimizing the overall operating cost.

Images