Industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning
By employing data decomposition and ensemble learning methods, the problems of data fluctuations and long model training cycles in load and photovoltaic forecasts in industrial and commercial energy storage systems have been solved, enabling efficient intelligent energy storage scheduling and improving forecast accuracy and profitability.
Patent Information
- Application Number
- CN202511424427.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-30
- Publication Date
- 2026-01-06
AI Technical Summary
Load forecasting and photovoltaic forecasting in industrial and commercial energy storage systems suffer from problems such as large data fluctuations, long model training cycles, difficulty in quickly capturing data changes, and difficulty in balancing safety and economy.
The load data is decomposed into seasonal, trend and residual terms using data decomposition technology. Combined with outlier detection and ensemble learning algorithms, the load and photovoltaic forecasts are performed using XGBoost and LightGBM models. The peak-valley price difference revenue maximization model is constructed by combining energy storage operation parameters for intelligent scheduling.
It improves the accuracy of load and photovoltaic forecasting, achieving a balance between safety, stability and economy, and increasing daily revenue by 28%.
Smart Images

Figure CN121282958A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical fields of virtual power plants, smart energy management, and energy storage technology, and in particular to a smart scheduling method for industrial and commercial energy storage based on data decomposition and integrated learning. Background Technology
[0002] Load forecasting is a technique for predicting electricity demand at a specific point in the future or over a specific period. It is typically used in power system dispatching and planning. Power load forecasting helps power companies better dispatch generating units, avoid power shortages or surpluses, and ensure the stable operation of the power system.
[0003] Photovoltaic forecasting refers to the technology of predicting the power generation of photovoltaic (PV) systems over a future period. Because PV power generation is affected by factors such as weather, seasons, and sunshine, PV forecasting has become particularly important, especially given the increasing penetration rate of PV systems.
[0004] Load forecasting and photovoltaic forecasting are major challenges for industrial and commercial energy storage systems, and accurate forecasting is crucial.
[0005] On the one hand, load data is greatly influenced by subjective human factors, leading to significant fluctuations. In existing technologies, many researchers do not conduct thorough analysis of load data, instead relying excessively on models. For example, the paper "A Day-ahead Electricity Load Forecast Based on an EMD-MLP Combined Model"... [1] Trivedi A uses the PatchTST model for day-ahead load forecasting. [2] They use ensemble models or deep neural network models to allow the model to learn data features on its own. However, for complex data, a particularly complex model is needed to learn the data features, which will greatly increase the training cycle of the model.
[0006] On the other hand, load data is affected by sudden behavioral changes (e.g., work stoppages and starts), which has a significant impact on day-ahead load forecasting, making it difficult for models to quickly capture these changes. Current technologies mostly employ outlier detection algorithms, such as clustering algorithms, but these often result in missed or over-detection.
[0007] Scheduling algorithms are a core issue for commercial and industrial energy storage systems. Currently, my country's commercial and industrial energy storage faces challenges in four areas: safety, economy, policy, and talent development. Among these, safety and economy are the two most critical. In existing technologies, many only consider the economic aspect, often neglecting the importance of safety. Summary of the Invention
[0008] To address the aforementioned technical challenges, this invention proposes an intelligent scheduling method for industrial and commercial energy storage based on data decomposition technology and integrated learning. This method is based on load forecasting, photovoltaic forecasting, and scheduling algorithms to formulate efficient charging and discharging strategies for industrial and commercial energy storage, taking into account safety, stability, and economy. Furthermore, it features rolling forecasting and real-time updates of the forecast results.
[0009] The technical solution adopted by this invention to solve its technical problem is: A smart scheduling method for industrial and commercial energy storage based on data decomposition and ensemble learning includes the following steps: Step S1. Data preprocessing: First, based on the user-side meter data and the energy storage-side meter data, calculate the user's actual load data, and then fill in missing values and filter it.
[0010] Furthermore, the missing value filling adopts a full-day adaptive length sliding window + bidirectional average (MA5) missing value filling, and performs bidirectional moving average on the daily load data.
[0011] Furthermore, the filtering employs Savitzky-Golay filtering, which uses convolution to fit a continuous subset of adjacent data points to a low-order polynomial using linear least squares.
[0012] Step S2. Seasonal decomposition: The load data from step S1 is decomposed into seasonal, trend, and residual terms using Statsmodels and an additive model. Then, polynomial fitting is performed on the trend term and low-pass filtering is performed on the residual term to obtain the final seasonal, trend, and residual terms.
[0013] Step S3. Outlier Detection: The outlier detection algorithm is used on the load data after step S1. It includes two parts: outlier waveform detection and outlier detection. The outlier waveform detection uses a supervised algorithm, and the outlier detection uses an unsupervised algorithm.
[0014] Furthermore, the abnormal waveform detection first involves labeling the load data for abnormal waveforms. The labeling principle is based on the complete waveforms of each day, with anomalies labeled at shutdown and start points. Each waveform is then classified into different levels of disorder. Finally, the ResNet algorithm is used to detect abnormal waveforms and obtain the detection results for each waveform.
[0015] Furthermore, the outlier detection is based on the detection results of each waveform. It mainly adopts an unsupervised algorithm to locate the outlier point of each abnormal waveform and obtain the final outlier detection result, which is divided into normal point, stop point and start point.
[0016] Step S4. Model Training and Testing: Based on the outlier detection results in Step S3 and the decomposed data in Step S2, feature construction, feature selection, and parameter tuning are performed sequentially. Then, five-fold cross-fusion training is conducted using XGBoost and LightGBM. Next, the model is tested according to the load prediction evaluation index. Finally, it is deployed using cloud-edge to obtain the final load prediction value.
[0017] Furthermore, the load forecasting evaluation indicators include average absolute percentage error, average absolute error, percentage of charging period difference greater than 0, and percentage of discharging period difference less than 10% of transformer capacity or maximum demand.
[0018] Furthermore, the cloud-edge deployment uses a five-fold cross-fusion approach for XGBoost and LightGBM, employing cloud deployment for XGBoost and LightGBM, while edge deployment is used for abnormal waveform detection.
[0019] Step S5. Photovoltaic Forecast: First, the photovoltaic data is seasonally decomposed. Then, the weather data is obtained by calling the Open-Meteo Weather API. Next, XGBoost and LightGBM are cross-fused and tested according to the photovoltaic forecast evaluation index. Finally, the data is deployed using cloud-edge-device architecture to obtain the final photovoltaic forecast value.
[0020] Step S6. Scheduling Algorithm: Based on the load forecast value from step S4 and the photovoltaic forecast value from step S5, plus the energy storage operation parameters, a peak-valley price difference revenue maximization model is constructed. The model is solved using the MindOpt optimizer, and the solution is then corrected. The model is tested according to the scheduling algorithm evaluation index to obtain the final energy storage charging and discharging strategy.
[0021] Furthermore, the energy storage operating parameters are divided into offline parameters and real-time parameters. Offline parameters include user net load demand, electricity price, energy storage charging efficiency, energy storage discharging efficiency, energy storage installed power, energy storage installed capacity, transformer capacity, minimum load for reverse current protection, and maximum load for demand. Real-time parameters include real-time SOC of the battery cluster, real-time SOH of the battery cluster, real-time charge / discharge protection status of the battery cluster, real-time minimum voltage of the battery cluster, and real-time maximum voltage of the battery cluster.
[0022] Furthermore, the rules for the aforementioned modification include six aspects: stabilization of energy storage charging and discharging power, stabilization of energy storage charging and discharging voltage, energy storage charging and discharging protection constraints, anti-reverse current constraints, overload control, and default value scheduling strategy.
[0023] Furthermore, the evaluation metrics for the scheduling algorithm include safety metrics, stability metrics, and economic metrics. Safety metrics include energy storage backflow prevention, demand management, capacity management, and energy storage charging and discharging protection mechanisms; stability metrics include stable energy storage charging and discharging power, stable energy storage voltage changes, and stable optimization algorithms; economic metrics include maximizing the revenue from peak-valley price differences, such as average daily revenue and average daily growth rate.
[0024] The beneficial effects of this invention are This invention is based on data decomposition technology, which fully analyzes the data itself and decomposes it into seasonal, trend, and residual terms to simplify the data. Then, it uses an outlier detection algorithm to deal with sudden changes in user behavior (e.g., shutdown and start-up), thereby greatly increasing the accuracy of prediction. Finally, based on load forecasts, photovoltaic forecasts, and energy storage operation parameters, it formulates efficient charging and discharging strategies for industrial and commercial energy storage, taking into account safety, stability, and economy.
[0025] This invention uses rolling predictions every 15 minutes to update the predictions in real time. In implementation cases, tests at multiple sites have shown that the average daily revenue based on this invention has increased by 28%. Attached Figure Description
[0026] Figure 1 This is a flowchart of the intelligent scheduling process for industrial and commercial energy storage in this invention.
[0027] Figure 2 This is a load forecasting architecture diagram.
[0028] Figure 3 This is a data annotation and display chart.
[0029] Figure 4 This is a diagram of the scheduling algorithm architecture.
[0030] Figure 5 This is a comparison chart of the average daily returns of manual charging / discharging strategies and flexible charging / discharging strategies.
[0031] Figure 6 This is a comparison chart of the average daily revenue of a stable operating site. Detailed Implementation
[0032] This invention is part of the virtual power plant concept and mainly relates to the technical fields of smart grids, intelligent energy management systems, energy storage technology, big data and ensemble learning, and scheduling algorithms. Specifically, big data and ensemble learning mainly include load forecasting, photovoltaic forecasting, and outlier detection; scheduling algorithms mainly include linear programming, nonlinear programming, and mixed-integer programming.
[0033] refer to Figure 1 In its specific implementation, the present invention mainly includes the following steps: Step S1. Data preprocessing: First, the load data is filled with missing values and filtered.
[0034] The missing value imputation uses a full-day adaptive length sliding window plus a two-way average (MA5) missing value imputation. If the proportion of missing values is greater than the threshold, data from the previous week's same day will be used for imputation, and so on; if there are missing values but they are less than the threshold, a full-day adaptive length sliding window plus a two-way average (MA5) missing value imputation is used, which can fully consider the characteristics of the data itself.
[0035] The filtering method used is Savitzky-Golay filtering. The data after missing value imputation is then subjected to Savitzky-Golay filtering to obtain the filtered data.
[0036] Step S2. Seasonal decomposition: Because load data is greatly influenced by subjective human factors, it fluctuates significantly. To fully extract the characteristics of the data, a data decomposition technique, namely seasonal decomposition, is used.
[0037] The load data is decomposed into seasonal, trend, and residual terms. Polynomial fitting is performed on the trend term, and low-pass filtering is applied to the residual term.
[0038] Specifically, the preprocessed data is used to decompose the load data into seasonal, trend, and residual terms using Statsmodels and an additive model. Then, polynomial fitting is performed on the trend term and low-pass filtering is applied to the residual term to obtain the final seasonal, trend, and residual terms.
[0039] Step S3. Outlier Detection: Because load data is affected by sudden changes in behavior, such as shutdowns and starts, this has a significant impact on day-ahead load forecasting, and models have difficulty quickly capturing these changes.
[0040] This invention uses an outlier detection algorithm, which includes two parts: outlier waveform detection and outlier detection. The outlier waveform detection uses a classification algorithm, while the outlier detection uses an unsupervised algorithm.
[0041] For abnormal waveform detection algorithms, the first step is to label the load data with abnormal waveforms. The labeling principle is based on the complete waveforms of each day, and anomalies are labeled at the stop and start points. Each waveform is also classified into different levels of disorder. Then, the ResNet algorithm is used to detect abnormal waveforms and obtain the detection results for each waveform.
[0042] Based on the detection results of each waveform, outlier detection is performed, mainly using unsupervised algorithms to locate outlier points in each abnormal waveform, resulting in the final outlier detection results, which are divided into normal points, stop points, and start points.
[0043] Step S4. Model Training and Testing: First, based on the outlier detection results, the data is categorized into normal points, work stoppage points, and work resumption points. The historical data obtained will differ depending on the outcome. For normal points, historical data will be obtained chronologically; for work stoppage points, historical holiday data will be obtained; and for work resumption points, historical weekday data will be obtained.
[0044] Secondly, based on historical data, and by seasonally decomposing it, we obtain seasonal terms, trend terms, and residual terms.
[0045] Finally, feature construction, feature selection, and parameter tuning will be performed on these three items respectively, and then XGBoost and LightGBM will be used for cross-fusion to obtain the final load prediction value.
[0046] Step S5. Photovoltaic Forecasting: Because photovoltaic data is greatly affected by weather factors, and because photovoltaic data also exhibits strong regularity, Therefore, the photovoltaic data is seasonally decomposed, and then weather data is used to assist in the cross-fusion using XGBoost and LightGBM.
[0047] Specifically, Statsmodels is first used to decompose photovoltaic data into seasonal, trend, and residual terms based on an additive model. Then, weather data is obtained by scheduling the Open-Meteo weather API. Based on these three terms, feature construction, feature selection, and parameter tuning are performed respectively. Finally, XGBoost and LightGBM are used for cross-fusion to obtain the final photovoltaic prediction value.
[0048] Step S6. Scheduling Algorithm: Based on load forecasts and photovoltaic forecasts, along with energy storage operation parameters, a peak-valley price difference revenue maximization model is constructed. The model is solved using the MindOpt optimizer, and the results are then corrected to obtain the final energy storage charging and discharging strategy.
[0049] Specifically, the difference between the load forecast and the photovoltaic forecast is first calculated. Then, based on energy storage operating parameters, such as SOC and energy storage charge / discharge protection status, an optimization model that maximizes the peak-valley price difference revenue is constructed. The model is solved using the MindOpt optimizer. Finally, the optimization results are corrected based on six aspects: energy storage charge / discharge power stabilization, energy storage charge / discharge voltage stabilization, energy storage charge / discharge protection constraints, anti-reverse current constraints, overload control, and default value scheduling strategy, to obtain the final energy storage charge / discharge strategy.
[0050] The core technologies of this invention mainly include load forecasting, photovoltaic forecasting, and scheduling algorithms. To more effectively evaluate algorithm performance, the following evaluation indicators are established:
[0051] Evaluation metrics for photovoltaic prediction algorithms: Traditional time series evaluation metrics: MAPE, MAE.
[0052] Scheduling algorithm evaluation metrics: Safety: Energy storage anti-backflow, demand management, capacity management, and energy storage charging and discharging protection mechanisms.
[0053] Stability: Energy storage charging and discharging power is stable, energy storage voltage changes are stable, and the optimization algorithm is stable.
[0054] Economic efficiency: Maximizing the profit from peak-valley price differences, such as average daily profit, average daily growth rate, etc.
[0055] The following is combined with Figure 2 and Figure 3 The load prediction of the present invention will be described in detail below: Depend on Figure 2 As can be seen, load forecasting mainly includes four steps: data preprocessing, seasonal decomposition, outlier detection, and model training and testing, which are steps S1-S4 in the technical solution of this invention.
[0056] Step S1: Data preprocessing. Data preprocessing mainly includes two aspects: missing value imputation and filtering.
[0057] Step S11 Missing value imputation: Missing value imputation uses a full-day adaptive length sliding window + bidirectional average (MA5) missing value imputation. If the proportion of missing values is greater than the threshold, data from the previous week's same day will be used for imputation, and so on; if there are missing values but they are less than the threshold, a full-day adaptive length sliding window + bidirectional average (MA5) missing value imputation will be used.
[0058] Step S12 Filtering: The filtering uses Savitzky-Golay filtering. The Savitzky-Golay filter is a digital filter that can be applied to a set of data to smooth it, improving data accuracy without altering the signal's trend or width. The Savitzky-Golay filter is implemented through convolution, that is, by fitting a continuous subset of adjacent data points to a low-order polynomial using linear least squares.
[0059] Advantages of the Savitzky-Golay filter: Different window widths can be selected at any position on the same curve to meet different smoothing filtering needs; it is especially advantageous for processing time-series data at different stages. The Savitzky-Golay filter is applied to the data after missing value imputation to obtain the filtered data.
[0060] Step S2 Seasonal Decomposition: Since the load data has certain regularity, it is seasonally decomposed. First, the preprocessed data is decomposed into seasonal, trend, and residual terms using Statsmodels and an additive model. Then, polynomial fitting is performed on the trend term and low-pass filtering is performed on the residual term to obtain the final seasonal, trend, and residual terms.
[0061] Step S3 Outlier Detection: Since load data is greatly affected by user subjective factors, the detection of shutdown and start-up points is crucial, as it will have the greatest impact on daytime load forecasting.
[0062] Outlier detection is divided into two parts: outlier waveform detection and outlier detection. Outlier waveform detection involves examining the waveforms of each day to determine if they are abnormal. However, the biggest challenge in outlier waveform detection is data annotation, which requires significant manpower and effort. Data annotation tools, such as… Figure 3 As shown, the main functions include waveform display (current waveform, waveform overlay, waveform template), waveform type, waveform disorder level, easily misjudged waveform type, previous waveform, next waveform, jump, backtracking, etc.; outlier detection is based on abnormal waveforms and performs outlier detection on them.
[0063] Step S31: Abnormal Waveform Detection: Abnormal waveforms are labeled on the load data. The labeling principle is based on the complete waveforms of each day, with anomalies labeled at shutdown and start-up points. Each waveform is then classified into different levels of disorder. The ResNet algorithm is then used for abnormal waveform detection to obtain the detection results for each waveform.
[0064] Step S32 Outlier Detection: Based on the detection results of each waveform, outlier detection is performed. This mainly uses an unsupervised algorithm to locate outlier points in each outlier waveform and obtain the final outlier detection results.
[0065] Step S4 Model Training and Testing: Model training and testing mainly includes six aspects: model selection, feature construction, feature selection, parameter tuning, training and testing, and deployment.
[0066] Step S41: The model is selected to primarily sample XGBoost and LightGBM for model fusion.
[0067] Step S42, feature construction, is divided into site features, time features, and historical features. The time features mainly consist of six features: year, quarter, day, week, hour, and minute. The historical features are the load data for the past two days, with a sampling frequency of 15 minutes.
[0068] Step S43 Feature Selection Sampling XGBoost and LightGBM models respectively obtain the score of each feature, then select the first 96 features of the two models, take their union, and obtain the final features.
[0069] Step S44 parameter tuning mainly involves Bayesian optimization.
[0070] Step S45: Training and Testing. Based on the selected optimal parameters and features, perform five-fold cross-training using XGBoost and LightGBM, and then test the data using load prediction evaluation metrics.
[0071] Step S46 Deployment: Deployment is mainly divided into a combination of cloud deployment and edge deployment. The advantages of cloud deployment include high performance and ease of deployment; the advantages of edge deployment include high timeliness and fast data transmission. Therefore, cloud deployment is mainly for tasks with low timeliness, such as load forecasting, photovoltaic forecasting, and scheduling algorithms; edge deployment is mainly for tasks with high timeliness, such as outlier detection and correction algorithms.
[0072] The photovoltaic forecast is explained in detail below: Because photovoltaic (PV) data is significantly affected by weather factors, accurate acquisition of local weather data is crucial. The PV forecasting process is similar to load forecasting, the only difference being feature construction. PV forecasting feature construction is divided into site features, time features, weather features, and historical features. Weather features include air pressure, relative humidity (2m), cloud cover, 10m wind speed, 10m wind direction, temperature (2m), radiation intensity (J / m²), precipitation, 100m wind speed, 100m wind direction, whether it is daytime or nighttime, sunshine duration, radiation intensity (J / m²)_max, temperature (K)_max, temperature (K)_min, 10m wind speed sin, 100m wind speed sin, first-order difference of radiation intensity, and eighth-order difference of radiation intensity. Historical features are PV data from the past two days, sampled at a frequency of 15 minutes. Then, feature selection, parameter tuning, model training, and testing are performed to obtain the final PV forecast value.
[0073] The following is combined with Figure 4 The scheduling algorithm of this invention is described in detail below: refer to Figure 4 The scheduling algorithm mainly consists of three parts: data preprocessing, scheduling algorithm, and correction algorithm. It is assumed that photovoltaic power is directly connected to the user side for priority use, with surplus power fed into the grid.
[0074] Step S61 Data Preprocessing: Calculate the power data involved in the scheduling, i.e., load forecast value minus photovoltaic forecast value.
[0075] Step S62 Scheduling Algorithm: The scheduling algorithm mainly includes five aspects: objective function construction, constraint selection, decision variable selection, offline and real-time variable selection, and optimization algorithm selection.
[0076] The decision variables include energy storage charging strategy, energy storage discharging strategy, energy storage energy state, energy storage charging state, and energy storage discharging state.
[0077] Offline parameters include user net load demand (the difference between load and photovoltaic power), electricity price, energy storage charging efficiency, energy storage discharging efficiency, energy storage installed power, energy storage installed capacity, transformer capacity, minimum load for backflow prevention, and maximum load for demand.
[0078] Real-time parameters include real-time SOC of battery cluster, real-time SOH of battery cluster, real-time charge / discharge protection status of battery cluster, real-time minimum voltage of battery cluster, and real-time maximum voltage of battery cluster.
[0079] Step S63 Correction Algorithm: The correction algorithm mainly includes six aspects: energy storage charging and discharging power stabilization, energy storage charging and discharging voltage stabilization, energy storage charging and discharging protection constraints, anti-reverse current constraints, overload control, and default value scheduling strategy.
[0080] Energy storage charging and discharging power stabilization. The strategy for each charging and discharging period is stabilized to make energy storage charging and discharging more stable.
[0081] Energy storage charging peak voltage stabilization: Obtain the peak voltage of the battery cluster at the current moment and the previous moment, calculate the difference between the two, and when the difference is greater than 0.025, reduce the charging power at the current moment.
[0082] Energy storage discharge minimum voltage stabilization: Obtain the minimum voltage of the battery cluster at the current moment and the previous moment, calculate the difference between the two, and when the difference is greater than 0.025, reduce the discharge power at the current moment.
[0083] Overload control and anti-reverse current constraint: Obtain the average load value of the previous five minutes and the energy storage charging and discharging strategy at the current moment. If it is a charging strategy, determine whether the average load value + charging strategy is greater than the transformer capacity. If it is, reduce the charging strategy. If it is a discharging strategy, determine whether the discharging strategy - average load value is greater than -10. If it is, reduce the discharging strategy.
[0084] Default scheduling strategy: Main reasons for missing scheduling strategy: 1. Load forecasting failure, causing scheduling to crash; 2. Constraint parameters in scheduling exceed thresholds, for example, load forecast values are too high, exceeding transformer capacity, leading to unsolvable situations; 3. Failure to acquire some real-time parameters in scheduling, causing scheduling to crash. Solution: Add a default scheduling strategy. Since the scheduling strategy varies significantly from site to site each month, manually setting a default scheduling strategy would be extremely time-consuming and labor-intensive. At 23:56 on the last day of each month, obtain the latest time-of-use electricity price for the following month and the actual load power of the previous week. Schedule the actual load power of the current week according to the latest time-of-use electricity price, calculate the average charging and discharging strategy for each time period of the day, and this becomes the default scheduling strategy.
[0085] The technical approach of this invention involves first preprocessing the load data and photovoltaic data, then training and testing the model, and finally deploying it. See details below. Figure 2 Then, based on the difference between the load forecast and the photovoltaic forecast, the forecast value participating in the dispatch is calculated. The dispatch and adjustment algorithms are then applied to optimize and correct this value, and finally, it is sent to the smart energy management system. See details... Figure 4 . Example
[0086] For day-ahead load forecasting, this invention first applies a full-day adaptive length sliding window with bidirectional average (MA5) missing value imputation and Savitzky-Golay filtering to the load data. Then, it performs seasonal decomposition based on the factory's working habits. Next, it performs feature selection and parameter tuning. Finally, it uses XGBoost and LightGBM models for five-fold cross-fusion.
[0087] The following test results were obtained based on the load forecasting evaluation indicators, and a comparative analysis was conducted with the PatchTST model. The test data period was from September 7, 2024 to October 7, 2024, with 107 test data sites, and the forecasting task was day-ahead forecasting (24 hours).
[0088] surface The table shows the test results of the load forecasting model and the PatchTST model of this invention on all load data, respectively. The table shows the test results of the load forecasting model and the PatchTST model of this invention on weekday load data, respectively. The table above shows the test results of the load forecasting model and the PatchTST model of this invention on non-working day load data. It can be seen that, for the three indicators MAPE, MAE, and Discharge, the load forecasting algorithm of this invention is superior to the PatchTST algorithm; for the Charge indicator, the load forecasting algorithm of this invention is slightly inferior to the PatchTST algorithm. In summary, the load forecasting algorithm of this invention is superior to the PatchTST algorithm.
[0089] For day-ahead photovoltaic (PV) forecasting, this invention first applies a full-day adaptive-length sliding window with bidirectional average (MA5) missing value imputation and Savitzky-Golay filtering to the PV data. Then, it performs seasonal decomposition, feature selection, and parameter tuning. Finally, it uses XGBoost and LightGBM models for cross-fusion. The test results are obtained based on PV forecasting evaluation metrics and compared with the PatchTST model. The test data period is from September 7, 2024 to October 7, 2024, with 17 test data sites, and the forecasting task is day-ahead forecasting (24 hours).
[0090] For the daytime scheduling algorithm, this invention first calculates the difference between the load forecast and the photovoltaic forecast to obtain the data participating in the scheduling. Then, it applies optimization and correction algorithms to calculate the average daily revenue from the peak-valley price difference and compares it with the manual charging and discharging strategy. The test data period is from October 30, 2024 to November 13, 2024, with 43 test data sites, and the scheduling task is daytime scheduling (24 hours).
[0091] Depend on Figure 5 It can be seen that the average daily growth rate of the manual charging and discharging strategy and the flexible charging and discharging strategy for 43 sites reached 47%, with the average daily revenue of the manual charging and discharging strategy being 433.74 yuan and the average daily revenue of the flexible charging and discharging strategy being 638.87 yuan.
[0092] Depend on Figure 6 It can be seen that, comparing the manual charging and discharging strategy and the flexible charging and discharging strategy for the 9 stable operating sites, the average daily growth rate reached 28%, with the average daily revenue of the manual charging and discharging strategy being 472.81 yuan and the average daily revenue of the flexible charging and discharging strategy being 605.29 yuan.
[0093] References
[0094] [1] Liu Luyao, Chen Zhigang, Shen Xinwei, et al. Day-ahead forecast of electricity load based on EMD-MLP combined model [J]. Southern Energy Construction, 2024, 11(01):143-156. DOI:10.16516 / j.ceec.2024.1.15. [2]Trivedi A. Explainable Forecasting Models For Load Forecasting[D]. University of Twente, 2024.
Claims
1. A commercial and industrial energy storage intelligent scheduling method based on data decomposition and ensemble learning, characterized in that, Comprise the following steps: Step S1. Data preprocessing: first, according to the user side electric meter data and energy storage side electric meter data, the actual load data of the user is calculated, and then the missing value filling and filtering are carried out; Step S2. Seasonal decomposition: the load data after step S1 is decomposed into seasonal term, trend term and residual term by using Statsmodels and based on additive model, then the trend term is fitted by polynomial and the residual term is low-pass filtered, so as to obtain the final seasonal term, trend term and residual term; Step S3. Abnormal value detection: the load data after step S1 is detected by using abnormal value detection algorithm, including abnormal waveform detection and abnormal value detection two parts, and the abnormal waveform detection adopts supervised algorithm, and the abnormal value detection adopts unsupervised algorithm; Step S4. Model training and testing: based on the abnormal value detection result of step S3 and the decomposed data of step S2, the feature construction, feature selection and parameter adjustment are carried out in turn, then XGBoost and LightGBM are used for five-fold cross fusion training, then the load prediction evaluation index is used for testing, finally the cloud-edge end is used for deployment, and the final load prediction value is obtained; Step S5. Photovoltaic prediction: first, the photovoltaic data is seasonally decomposed, then the weather data is obtained by calling Open-Meteo weather API, then XGBoost and LightGBM are used for cross fusion, and the photovoltaic prediction evaluation index is used for testing, finally the cloud-edge end is used for deployment, and the final photovoltaic prediction value is obtained; Step S6. Dispatching algorithm: based on the load prediction value of step S4 and the photovoltaic prediction value of step S5, the peak-valley price difference income maximization model is constructed by adding the energy storage operation parameters, the MindOpt optimizer is used for solving, then the solving result is corrected, and the dispatching algorithm evaluation index is used for testing, and the final energy storage charging and discharging strategy is obtained.
2. The industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning according to claim 1, characterized in that, The missing value filling in step S1 adopts full-time adaptive length sliding window+bidirectional average (MA5) missing value filling, and the bidirectional moving average is carried out on the load data of each day.
3. The industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning according to claim 1, characterized in that, The filtering in step S1 adopts Savitzky-Golay filtering, and the Savitzky-Golay filter is fitted with a low-order polynomial by convolution using linear least squares method.
4. The industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning according to claim 1, characterized in that, The abnormal waveform detection in step S3 first carries out data labeling of abnormal waveform of load data, and the data labeling principle will be based on the complete waveform of each day, and the abnormal points of shutdown point and startup point are labeled, and the waveform disorder level of each waveform is divided, then the ResNet algorithm is used for abnormal waveform detection, and the detection result of each waveform is obtained.
5. The industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning according to claim 1, characterized in that, The abnormal value detection in step S3 is based on the detection result of each waveform, and the abnormal value detection is carried out, mainly using unsupervised algorithm, locating the abnormal points of each abnormal waveform, and obtaining the final abnormal value detection result, which is divided into normal points, shutdown points and startup points.
6. The industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning according to claim 1, characterized in that, The load prediction evaluation index in step S4 includes the mean absolute percentage error, the mean absolute error, the proportion of charging period difference greater than 0, and the proportion of discharging period difference less than 10% of the transformer capacity or the maximum demand.
7. The industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning according to claim 1, characterized in that, In step S4, the cloud-edge deployment is performed on XGBoost and LightGBM using five-fold cross-validation fusion and cloud deployment for abnormal waveform detection using edge deployment. 8.The industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning of claim 1, characterized in that, The energy storage operation parameters in step S6 are divided into offline parameters and real-time parameters. The offline parameters include user net load demand, electricity price, energy storage charging efficiency, energy storage discharging efficiency, energy storage installed power, energy storage installed capacity, transformer capacity, anti-backflow minimum load, and demand maximum load. The real-time parameters include battery cluster real-time SOC, battery cluster real-time SOH, battery cluster real-time charging and discharging protection state, battery cluster real-time minimum voltage, and battery cluster real-time maximum voltage.
9. The industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning according to claim 1, characterized in that, The correction in step S6 includes six aspects: energy storage charging and discharging power smoothing, energy storage charging and discharging voltage smoothing, energy storage charging and discharging protection constraint, anti-backflow constraint, excess control, and default value scheduling strategy.
10. The industrial and commercial energy storage intelligent scheduling method based on data decomposition and ensemble learning according to claim 1, characterized in that, The scheduling algorithm evaluation index in step S6 includes safety index, stability index, and economic index. The safety index includes energy storage anti-backflow, demand management, capacity management, and energy storage charging and discharging protection mechanism. The stability index includes energy storage charging and discharging power stability, energy storage voltage change stability, and optimization algorithm stability. The economic index includes the maximum profit of peak-valley price difference, such as daily average profit and daily average growth rate.
