Load short-term prediction method considering optical storage unit based on data-knowledge fusion driving
By using a data-knowledge fusion-driven approach, combining the interactive response knowledge of photovoltaic energy storage units with common prediction factors, the problem of low load prediction accuracy and poor generalization in existing technologies is solved, achieving higher accuracy and better generalization in load prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-03-13
AI Technical Summary
Existing model-driven and data-driven load forecasting methods struggle to fully consider complex factors such as joint photovoltaic-storage scheduling and demand response when considering user-side photovoltaic energy storage units, resulting in low forecast accuracy, poor generalization performance, and insufficient interpretability.
A data-knowledge fusion-driven approach is adopted, which uses the interactive response knowledge of photovoltaic energy storage units and common predictive influencing factors as input features to construct a data-knowledge fusion-driven model. This model combines photovoltaic-storage response signals, user response rates and their uncertainties to perform load forecasting.
It improves the accuracy and generalization of load forecasting, reduces dependence on datasets, and achieves forecast results that are more closely aligned with users' actual response levels.
Smart Images
Figure CN121663466A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power system operation, and more specifically relates to a short-term load forecasting method based on data-knowledge fusion and considering photovoltaic-storage units. Background Technology
[0002] As the penetration rate of renewable energy sources such as solar and wind power continues to increase, the problems caused by their intermittent nature are becoming increasingly prominent. Because renewable energy sources are greatly affected by natural weather factors, they struggle to provide stable power, posing significant challenges to grid voltage, frequency, generation dispatch, and reliability. The emergence of energy storage technology has effectively solved this problem. Its joint dispatch with renewable energy sources for grid voltage regulation, frequency regulation, and peak shaving is of great significance for maintaining grid stability, improving power quality, and achieving supply-demand balance.
[0003] Furthermore, with the widespread implementation of demand response on the distribution network side, the charging and discharging behavior of user-side energy storage devices is constrained through electricity price incentives and intelligent regulation, in conjunction with rooftop photovoltaic power generation, thereby improving energy utilization and smoothing the load curve. The joint dispatch of photovoltaic and energy storage, coupled with the impact of demand response, alters load characteristics and increases load complexity. Therefore, when conducting short-term load forecasting considering photovoltaic and energy storage units, it is necessary to incorporate the joint dispatch model of photovoltaic and energy storage, along with the demand response adjustment effect, into the forecasting model to obtain more accurate prediction results.
[0004] For short-term load forecasting considering user-side self-consumption photovoltaic and energy storage equipment, it is necessary to consider not only constraints such as meteorological factors and equipment physical characteristics, but also the impact of joint scheduling and demand response factors. Therefore, in this scenario, traditional load forecasting methods based on model-driven forecasting and data-driven forecasting have significant limitations: model-driven forecasting methods can consider very limited influencing factors, are difficult to model, and have poor generalization performance; data-driven forecasting methods rely on accurate and complete datasets, and artificial intelligence-based data-driven methods have weak interpretability.
[0005] Therefore, considering the load forecasting scenario of user-side photovoltaic energy storage units, taking into account the correlation between load and complex influencing factors, joint scheduling, demand response, and other characteristics are technical problems that urgently need to be solved. Summary of the Invention
[0006] The purpose of this invention is to provide a short-term load forecasting method based on data-knowledge fusion and considering photovoltaic-storage units. This method achieves load forecasting with moderate modeling difficulty, high prediction accuracy, good generalization, and strong interpretability.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is: a short-term load forecasting method based on data-knowledge fusion considering photovoltaic-storage units, characterized by the following steps: The interactive response knowledge of photovoltaic energy storage units and common predictive influencing factors are used as input features and introduced into the data-driven model in parallel to construct a data-knowledge fusion-driven model. The model is then trained to obtain load prediction values with high prediction accuracy that are close to the actual response level of users. The photovoltaic energy storage unit interactive response knowledge includes: the photovoltaic energy storage response signal extracted from the photovoltaic energy storage unit joint scheduling model, and the user response rate and its uncertainty under different grid-connected electricity price levels; The common factors affecting forecasting include: electricity price, temperature, humidity, and historical load data; The data-knowledge fusion-driven model uses knowledge such as the optical storage response signal and the user response rate taking into account uncertainty, along with other common prediction influencing factors, as input features to establish a data-knowledge fusion-driven prediction model, and trains the model under the dual drive of knowledge and data.
[0008] Furthermore, the photovoltaic-storage response signal is obtained by solving the photovoltaic-storage unit joint scheduling model; the photovoltaic-storage unit joint scheduling model includes: energy storage unit charging and discharging power model, energy storage unit charging and discharging constraints, and energy storage device physical characteristics.
[0009] Furthermore, when the photovoltaic power generation of the photovoltaic energy storage unit exceeds the load demand or the electricity price during that period is a low-off-peak price, the energy storage unit charges; when the photovoltaic power generation is less than the load demand and the electricity price during that period is a peak price, the energy storage unit discharges. This is used to build a charging and discharging power model for the energy storage unit. in and For energy storage units t Charging and discharging power during the same period; and They are respectively t The excess photovoltaic power generation during a given period and the power purchased by users from the grid; This represents the maximum power output for each charge and discharge cycle of the energy storage device. for t The remaining energy storage capacity after the -1 period ends; and These are the minimum remaining capacity limit and the maximum remaining capacity limit for energy storage, respectively. and These represent the charging and discharging efficiencies of energy storage, respectively. and Let these represent the set of off-peak electricity prices and the set of peak electricity prices, respectively. For time intervals.
[0010] Furthermore, to ensure the normal operation of photovoltaic energy storage units and the stable operation of the power grid, it is stipulated that... t During a given period, a single energy storage device cannot simultaneously charge and discharge: in The variables are 0-1, where 0 indicates no charging / discharging behavior and 1 indicates charging / discharging behavior, thus constraining the individual energy storage unit in... t Charging and discharging do not occur simultaneously within the same time period; The value of the photovoltaic-storage response sub-signal of the photovoltaic-storage unit in time period t is the charging and discharging power of the energy storage unit. A positive value indicates energy storage discharge, and a negative value indicates energy storage charging. The sequence of optical storage response signals within a day; and For energy storage units t Charging and discharging power during the same period; This represents the demand response signal of the photovoltaic energy storage unit at time t, where (The remaining numbers 1, 2, T-1, T, etc. do not need to be explained in detail).
[0011] Furthermore, the charge and discharge characteristics of the energy storage device are modeled: in This refers to the self-discharge rate of the energy storage device.
[0012] Furthermore, the level of economic incentives, a common predictive factor, affects the charging and discharging power of photovoltaic energy storage; based on consumer psychology, the user response curve, taking uncertainty into account, is divided into dead zone, linear zone, and saturation zone; and optimistic scenarios are utilized. and pessimistic situations Describe the user response rate: In the formula p For economic incentives; and These represent the dead-zone response rate under optimistic and pessimistic scenarios, respectively. and These represent the slopes of the response rate curves under optimistic and pessimistic conditions, respectively. a and b These are critical excitation and saturation excitation, respectively. This represents the user's saturation response rate.
[0013] Furthermore, considering the uncertainty of user response rate, we assume that the user response rate follows a normally distributed random variable within an interval with upper and lower limits under optimistic and pessimistic scenarios. ; This represents the actual user response rate, where p is the economic incentive; N indicates a normal distribution. This represents the expected value of the actual user response rate under economic incentive p; This represents the variance of the actual user response rate under economic incentive p.
[0014] Furthermore, the common predictive influencing factors include: t The system considers electricity price, temperature, and humidity at specific times; it employs an iterative load forecasting method to predict the load at the next time point based on the historical load at the previous four time points.
[0015] Furthermore, the interactive response knowledge of photovoltaic energy storage units and common predictive influencing factors are used as input features to introduce a data-driven model. The data-driven model includes machine learning models such as LSTM (Long Short-Term Memory) neural networks, BP (Backpropagation) neural networks, and SVM (Support Vector Machine).
[0016] Furthermore, the prediction evaluation indicators selected are the mean square error (RMSE) and the mean relative percentage error (MAPE): In the formula, RMSE represents the mean square error; MAPE represents the mean relative percentage error; and n represents the total number of load samples. This represents the i-th load data; This represents the predicted value of the i-th load data. This represents the actual value of the i-th load data.
[0017] Furthermore, the proposed method is applied as follows: the trained data-knowledge fusion-driven prediction model is deployed to the distribution network side, and the photovoltaic-storage response signal, the user response rate considering uncertainty, and other prediction influencing factors are input into the model in parallel to obtain load prediction results with high prediction accuracy and more in line with the actual response level of users after considering the photovoltaic-storage interaction response factors, thus providing a data foundation for the optimized scheduling of the distribution network and power marketing.
[0018] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: This invention establishes a joint scheduling model for photovoltaic (PV) energy storage units, considering the impact of user-side PV power generation and energy storage charging / discharging on load changes, and extracts PV-energy storage response signals. This invention establishes a user response rate model that takes uncertainty into account, representing the user response rate through optimistic and pessimistic scenarios, and representing the uncertainty of the user response rate under a certain economic incentive as a normally distributed random variable, thus more closely reflecting the actual user response level. This invention incorporates two types of load knowledge—PV-energy storage response signals and user response rates and their uncertainties—along with common load influencing factors such as electricity prices, temperature and humidity, and historical load data, as features into a data-driven model, establishing a data-knowledge fusion-driven model. This improves the model's prediction accuracy and generalization ability, and reduces its dependence on datasets. Thus, this invention can consider the impact of interactive response factors such as the charging and discharging behavior of user-side PV energy storage units on load changes. Through the fusion of data and knowledge, it reduces modeling difficulty and achieves load forecasting with moderate modeling difficulty, high prediction accuracy, good generalization ability, and strong interpretability. Attached Figure Description
[0019] Figure 1 This is a framework diagram of a short-term load forecasting method for photovoltaic-storage units based on data-knowledge fusion, provided in an embodiment of the present invention.
[0020] Figure 2 This is a diagram illustrating the charging and discharging modes of the photovoltaic energy storage unit provided in an embodiment of the present invention.
[0021] Figure 3 This is a user response rate pattern diagram that takes uncertainty into account, provided for embodiments of the present invention.
[0022] Figure 4 A flowchart for load prediction of photovoltaic-storage units based on data-knowledge fusion is provided for embodiments of the present invention.
[0023] Figure 5 shows the prediction results for a specific day in the test set provided in this embodiment of the invention. (a) is the load prediction curve of the LSTM model, (b) is the load prediction curve of the BP model, and (c) is the load prediction curve of the SVM model. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0025] See Figure 1 This invention provides a method for short-term load forecasting of photovoltaic-storage units based on data-knowledge fusion, comprising: By combining two types of load knowledge—photovoltaic-storage response signals and user response rates that take uncertainty into account—with common predictive influencing factors such as electricity prices, temperature and humidity, and historical load data, and introducing them in parallel as input features into the data-driven model, a data-knowledge fusion-driven model is obtained. Specifically, the photovoltaic-storage response signal refers to the operating principle of "self-consumption and surplus power to the grid" in the joint dispatch model of photovoltaic energy storage units. t The 0-1 signals represent the charging and discharging behavior of the energy storage unit at any given time. The charging and discharging modes of the photovoltaic energy storage unit are as follows: Figure 2 As shown.
[0026] The user response rate and its uncertainty are described under a certain economic incentive level. Specifically, based on consumer psychology, response curves for optimistic and pessimistic scenarios are obtained to describe user response rates. The actual user response rate is a random variable following a normal distribution within an interval bounded by the optimistic and pessimistic scenarios, thus describing the uncertainty of the user response rate. The user response rate pattern considering uncertainty is as follows: Figure 3 As shown.
[0027] The data-knowledge fusion-driven model includes: knowledge extraction, knowledge input, and data-driven modeling. Specifically, the photovoltaic-storage unit joint scheduling model extracts the photovoltaic-storage response signal, and the user response rate and its uncertainty model extract the actual user response rate, thus completing knowledge extraction. These two types of load knowledge, along with common load forecasting influencing factors, are used as input features and input into the data-driven model in parallel, completing knowledge input. Appropriate hyperparameters are selected for model training, completing data-driven modeling. The specific forecasting process is as follows: Figure 4 As shown.
[0028] The above method will be explained in detail below using contract data from a certain market as an example. The dataset is divided into training set and test set according to a ratio of 0.9:0.1.
[0029] In the joint scheduling model of photovoltaic energy storage units, the maximum power of each charge and discharge cycle of energy storage is set. The minimum energy storage capacity is 20kW. and maximum capacity limit The charging and discharging power of the energy storage device is 15% and 90% of its rated capacity, respectively. and The self-discharge rates of energy storage devices are 95% and 96%, respectively. It is 0.01.
[0030] In the user response rate and its uncertainty model, a dead-zone response rate is defined under the optimistic scenario. The dead zone response rate is 10% under pessimistic conditions. -5%, user saturation response rate 95%, critical excitation a and saturation excitation b These are 1 times and 1.25 times the contracted electricity price, respectively.
[0031] To analyze the impact of two types of knowledge—photovoltaic-storage response signal and user response rate considering uncertainty—on load forecasting, and to explore the generalization of data-knowledge fusion-driven methods, three data-driven models were selected: Long Short-Term Memory (LSTM) neural network, Backpropagation (BP) neural network, and Support Vector Machine (SVM). Three forecasting scenarios were set up for each data model: Scenario 1: Without considering the optical storage response signal, user response rate and its uncertainty, predict the load at the next moment; Scenario 2: Considering the optical storage response signal, but not the user response rate and its uncertainty, predict the load at the next moment; Scenario 3: Considering the optical storage response signal, user response rate and its uncertainty, predict the load at the next moment.
[0032] The parameter settings for each data-driven model are as follows: LSTM model hyperparameters: number of hidden neurons: 50, batch size: 50, maximum number of iterations: 200, initial learning rate: 0.007, learning rate decreases to 0.0035 after 100 iterations. BP model hyperparameters: number of iterations: 200, error threshold: The learning rate is 0.01. SVM model hyperparameters: kernel type is Gaussian kernel, penalty factor is 10.
[0033] The mean square error (RMSE) and mean relative percentage error (MAPE) are used as prediction evaluation indicators.
[0034] The prediction results for a specific day in the test set were selected for comparative analysis, as shown in Figure 5.
[0035] The entire test set was selected, and the RMSE and MAPE of each model under different prediction scenarios were calculated. The results are shown in Table 1.
[0036] Table 1. RMSE and MAPE for different prediction scenarios considering photovoltaic-storage units. Among the LSTM, BP, and SVM models, incorporating the optical storage response signal reduces prediction error. Furthermore, adding the user response rate, which accounts for uncertainty, further reduces prediction error. Under the LSTM model, compared to scenario one, scenario two shows a reduction of RMSE of 0.8666kW, a reduction of MAPE of 0.0745%, and a 4.09% improvement in prediction accuracy. Scenario three, compared to scenario two, shows a reduction of RMSE of 0.0694kW, a reduction of MAPE of 0.0359%, and a further 2.05% improvement in prediction accuracy. Under the BP model, compared to scenario one, scenario two shows a reduction of RMSE of 0.4608kW, a reduction of MAPE of 0.0410%, and a 2.72% improvement in prediction accuracy. Scenario three, compared to scenario two, shows a reduction of RMSE of 0.7495kW, a reduction of MAPE of 0.1241%, and an 8.44% improvement in prediction accuracy. Under the SVM model, compared with scenario one, scenario two saw a reduction of 0.1957kW in RMSE, a reduction of 0.0422% in MAPE, and a 2.95% improvement in prediction accuracy; compared with scenario two, scenario three saw a reduction of 0.1786kW in RMSE, a reduction of 0.0374% in MAPE, and a further 2.69% improvement in prediction accuracy.
[0037] Based on the above prediction results, the following conclusions can be drawn: Under the same data-driven model and different prediction scenarios, adding knowledge of the photovoltaic-storage response signal can improve the prediction accuracy; on this basis, adding knowledge of the user response rate and its uncertainty can correct the load prediction curve and further improve the prediction accuracy.
[0038] Therefore, with the addition of load knowledge, the overall prediction error of the three data-driven models, namely LSTM neural network, BP neural network and support vector machine (SVM), all show a decreasing trend, indicating that the data-knowledge fusion-driven method has good generalization ability in heating-type electrical load prediction.
[0039] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A short-term load forecasting method for photovoltaic-storage units based on data-knowledge fusion, characterized in that... Includes the following steps: The interactive response knowledge of photovoltaic energy storage units and common predictive influencing factors are used as input features and introduced into the data-driven model in parallel to construct a data-knowledge fusion-driven model. The model is then trained to obtain load prediction values with high prediction accuracy that are close to the actual response level of users. The photovoltaic energy storage unit interactive response knowledge includes: the photovoltaic energy storage response signal extracted from the photovoltaic energy storage unit joint scheduling model, and the user response rate and its uncertainty under different grid-connected electricity price levels; The common factors affecting forecasting include: electricity price, temperature, humidity, and historical load data; The data-knowledge fusion-driven model uses the optical storage response signal and user response rate knowledge that takes uncertainty into account, along with other common predictive influencing factors, as input features to establish a data-knowledge fusion-driven prediction model, and trains the model under the dual drive of knowledge and data.
2. The method for short-term load forecasting based on data-knowledge fusion considering photovoltaic-storage units according to claim 1, characterized in that, The photovoltaic-storage response signal is obtained by solving the photovoltaic-storage unit joint scheduling model; the photovoltaic-storage unit joint scheduling model includes: energy storage unit charging and discharging power model, energy storage unit charging and discharging constraints, and energy storage equipment physical characteristics.
3. The method for short-term load forecasting based on data-knowledge fusion considering photovoltaic-storage units according to claim 1, characterized in that, When the photovoltaic power generation of the photovoltaic energy storage unit exceeds the load demand or the electricity price during that period is at a low off-peak price, the energy storage unit charges; when the photovoltaic power generation is less than the load demand and the electricity price during that period is at a peak price, the energy storage unit discharges. This is used to build a charging and discharging power model for the energy storage unit. in and For energy storage units t Charging and discharging power during the same period; and They are respectively t The excess photovoltaic power generation during a given period and the power purchased by users from the grid; This represents the maximum power output for each charge and discharge cycle of the energy storage device. for t The remaining energy storage capacity after the -1 period ends; and These are the minimum remaining capacity limit and the maximum remaining capacity limit for energy storage, respectively. and These represent the charging and discharging efficiencies of energy storage, respectively. and Let these represent the set of off-peak electricity prices and the set of peak electricity prices, respectively. For time intervals.
4. The method for short-term load forecasting based on data-knowledge fusion considering photovoltaic-storage units according to claim 1, characterized in that, To ensure the normal operation of photovoltaic energy storage units and the stable operation of the power grid, it is stipulated that... t During a given period, a single energy storage device cannot simultaneously charge and discharge: in The variables are 0-1, where 0 indicates no charging / discharging behavior and 1 indicates charging / discharging behavior, thus constraining the individual energy storage unit in... t Charging and discharging do not occur simultaneously within the same time period; The value of the photovoltaic-storage response sub-signal of the photovoltaic-storage unit in time period t is the charging and discharging power of the energy storage unit. A positive value indicates energy storage discharge, and a negative value indicates energy storage charging. The sequence of optical storage response signals within a day; and For energy storage units t Charging and discharging power during the same period; This represents the demand response signal of the photovoltaic energy storage unit at time t, where .
5. The method for short-term load forecasting based on data-knowledge fusion considering photovoltaic-storage units according to claim 1, characterized in that, Modeling the charge and discharge characteristics of energy storage devices: in This refers to the self-discharge rate of the energy storage device.
6. The method for short-term load forecasting based on data-knowledge fusion considering photovoltaic-storage units according to claim 1, characterized in that, Among common predictive factors, the level of economic incentives affects the charging and discharging power of photovoltaic energy storage; based on consumer psychology, the user response curve, taking uncertainty into account, is divided into dead zone, linear zone, and saturation zone; and optimistic scenarios are utilized. and pessimistic situations Describe the user response rate: In the formula p For economic incentives; and These represent the dead-zone response rate under optimistic and pessimistic scenarios, respectively. and These represent the slopes of the response rate curves under optimistic and pessimistic conditions, respectively. a and b These are critical excitation and saturation excitation, respectively. This represents the user's saturation response rate.
7. The method for short-term load forecasting based on data-knowledge fusion considering photovoltaic-storage units according to claim 1, characterized in that, Considering the uncertainty of user response rate, we assume that the user response rate follows a normal distribution random variable within an interval with upper and lower limits under optimistic and pessimistic scenarios. ; This represents the actual user response rate, where p is the economic incentive; N indicates a normal distribution. This represents the expected value of the actual user response rate under economic incentive p; This represents the variance of the actual user response rate under economic incentive p.
8. The method for short-term load forecasting based on data-knowledge fusion considering photovoltaic-storage units according to claim 1, characterized in that, The common predictive factors include: t The system considers electricity price, temperature, and humidity at specific times; it employs an iterative load forecasting method to predict the load at the next time point based on the historical load at the previous four time points.
9. The method for short-term load forecasting based on data-knowledge fusion considering photovoltaic-storage units according to claim 1, characterized in that, The interactive response knowledge of photovoltaic energy storage units and common predictive influencing factors are used as input features to introduce a data-driven model; the data-driven model includes: LSTM neural network, BP neural network and SVM machine learning model.
10. The method for short-term load forecasting based on data-knowledge fusion considering photovoltaic-storage units according to claim 1, characterized in that, The prediction evaluation indicators selected are the mean square error (RMSE) and the mean relative percentage error (MAPE). In the formula, RMSE represents the mean square error; MAPE represents the mean relative percentage error; and n represents the total number of load samples. This represents the i-th load data; This represents the predicted value of the i-th load data. This represents the actual value of the i-th load data.