A user net load interval prediction method based on physical law constraint containing distributed photovoltaic

Abstract

This invention belongs to the field of power load range prediction technology with distributed power sources, specifically a user net load range prediction method based on physical constraints for distributed photovoltaic (PV) systems. This method first performs non-intrusive decoupling driven by sunlight, breaking down the net load into independent theoretical estimates of actual user electricity consumption and PV output. Then, considering the distinct fluctuation characteristics of actual user electricity consumption and PV output, appropriate range prediction models are constructed for each. Finally, a dynamic fusion strategy guided by both quantitative correlation and physical laws is used to generate the net load range. This method balances range prediction accuracy and physical interpretability, providing reliable technical support for grid dispatching and energy storage optimization in high-proportion PV integration scenarios. This method uses sunlight data as the core driver to decouple PV output from actual user electricity consumption and generates accurate net load ranges through component-based range prediction and multi-dimensional probabilistic fusion.

Description

Technical Field

[0001] This invention belongs to the field of power load interval prediction technology with distributed power sources, specifically a user net load interval prediction method with distributed photovoltaic power sources based on physical law constraints. Background Technology

[0002] With the large-scale application of distributed photovoltaic (PV) power on the user side, net load (the difference between the user's actual electricity load and PV output) has become a key data point for power system operation and management. Accurate net load range forecasting can provide risk warnings and decision-making basis for the power system; however, current net load range forecasting faces three major problems:

[0003] 1) Lack of decoupling mechanism: Traditional methods often directly model the net load as a whole without separating the independent fluctuation characteristics of load and photovoltaic output. The two significantly different data are mixed and modeled, which cannot accurately reflect the weight of their respective impact on the net load, resulting in a large prediction deviation in the net load range.

[0004] 2) Simplistic distribution assumptions: The normal distribution assumption is commonly used to describe the fluctuation patterns of load and photovoltaic power. However, in actual scenarios, the load is affected by seasons and user behavior, resulting in a non-normal distribution. The photovoltaic output is affected by cloud cover and sudden changes in irradiance, resulting in a skewed distribution. The simple normal distribution assumption is out of touch with reality.

[0005] 3) Crude integration method: Even when attempting to predict the components separately, the method often adopts a simple superposition of "load range ± photovoltaic range", ignoring the correlation between the historical probability distributions of the two and seasonal differences, and failing to generate a net load range that conforms to physical laws.

[0006] In summary, existing technologies cannot solve the core problems of "inaccurate decoupling, unrealistic component forecasting, and simplistic interval fusion," and there is an urgent need for a net load interval forecasting method that is constrained by physical laws and conforms to the characteristics of actual data. Summary of the Invention

[0007] This invention addresses the problems of inaccurate decoupling, unrealistic component forecasting, and simplistic interval fusion in current net load range forecasting. It provides a user net load range forecasting method based on physical constraints, specifically for distributed photovoltaic (PV) systems. This method uses solar irradiance data as its core driver, decoupling PV output from actual user electricity load. Through component interval forecasting and multi-dimensional probability fusion, it generates accurate net load fluctuation ranges, suitable for large-scale distributed PV applications on the user side, providing data support for the safe and stable operation of the power system.

[0008] This invention is implemented using the following technical solution: a user net load range prediction method based on physical constraints for distributed photovoltaic (PV) systems. First, a non-intrusive decoupling of the load-driven process is performed, splitting the net load into independent theoretical estimates of the user's actual electricity load and PV output. Then, considering the distinct fluctuation characteristics of the user's actual electricity load and PV output, suitable range prediction models are constructed for each. On the load side, user behavior patterns are focused to obtain the load prediction range, while on the PV side, physical constraints are embedded to obtain the PV output prediction range. Finally, the load prediction range and the PV output prediction range are dynamically fused using a dual-guided strategy of quantitative correlation and physical laws to generate a net load prediction range that conforms to both statistical data characteristics and strictly adheres to physical laws. This method balances range prediction accuracy and physical interpretability, providing reliable technical support for grid dispatching and energy storage optimization in high-proportion PV integration scenarios.

[0009] The aforementioned method for predicting user net load intervals with distributed photovoltaic power generation based on physical constraints includes the following non-intrusive decoupling process:

[0010] Step 1: Data Acquisition and Preprocessing

[0011] Collect historical net load of user and grid connection points Irradiance Ambient temperature And the basic parameters of the photovoltaic system, including installed capacity. Rated conversion efficiency and temperature coefficient ;

[0012] Step 2: Decomposing the load and photovoltaic output

[0013] 2.1 Modeling of Theoretical Photovoltaic Output

[0014] Based on the IV characteristic curve of the photovoltaic system, the irradiance is fitted using the random forest algorithm. Ambient temperature Theoretical estimates of photovoltaic output The nonlinear relationship between them is used to construct a three-dimensional mapping model; the input features of the three-dimensional mapping model include irradiance. Ambient temperature and temperature-corrected conversion efficiency The output is a theoretical estimate of the photovoltaic power output. , ,in, In the formula, For reference temperature;

[0015] 2.2 Reverse splitting of load and photovoltaic output

[0016] Based on historical net load Theoretical estimates of photovoltaic power output Based on this, a dynamic correction term is introduced. After eliminating photovoltaic fluctuations, the actual electricity load of users is estimated by reverse derivation. , .

[0017] The aforementioned method for predicting user net load intervals with distributed photovoltaic power based on physical constraints further includes, in step two:

[0018] 2.3 Decoupling Quality Assessment and Decoupling Iteration

[0019] To evaluate the rationality of the non-invasive decoupling results, a dual evaluation mechanism is constructed.

[0020] 1) Physical rationality test

[0021] Decoupled user actual power load It conforms to physical laws: its value is non-negative for the vast majority of the time period; its instantaneous rate of change Within a reasonable range for the application power scenario, avoid physically impossible power surges; the load curve should exhibit a daily periodicity consistent with user behavior patterns;

[0022] 2) Residual sequence analysis

[0023] Reconstructing residuals by calculating net load Quantitatively assess the overall consistency of decoupling, and evaluate the net load reconstruction residual. Defined as: , the residual sequence mean with standard deviation Preset standards must be met;

[0024] If the above test fails, decoupling iterative optimization is triggered.

[0025] In the aforementioned method for predicting the net load interval of users with distributed photovoltaic power based on physical constraints, step 2.1 involves using a random forest algorithm to fit the theoretical estimate of photovoltaic power output. Compared with historical actual photovoltaic output To correct the deviation, optimize the model parameters and reduce the theoretical estimate of photovoltaic output. Compared with historical actual photovoltaic output To minimize deviations, ensure that the model fit is R² ≥ 0.9.

[0026] The specific process for constructing an adapted interval prediction model for the above-mentioned user net load interval prediction method based on physical law constraints is as follows:

[0027] Step 3: Build an interval prediction model

[0028] 3.1 Quantitative Division of Seasonal Zones

[0029] To refine the modeling of the differentiated impacts of climate and user behavior on load and photovoltaic output, the year is divided into several typical seasonal zones based on temperature and electricity consumption patterns: the main season is divided into winter, summer, and spring / autumn based on daily average temperature; each main season is further subdivided into peak electricity consumption sub-zones and stable electricity consumption sub-zones based on electricity consumption intensity.

[0030] 3.2 Construction of Load Interval Prediction Model

[0031] An XGBoost point prediction + temporal error filtering interval model is constructed to balance behavioral pattern capture and uncertainty quantification, as detailed below:

[0032] 1) Feature Engineering

[0033] Based on historical net load data Irradiance Ambient temperature Raw time-series data, filtered for multiple characteristic parameters highly correlated with electricity load:

[0034] Time features: Extract hour codes, date types, and seasonal partition identifiers from time-series data;

[0035] Environmental characteristics: average temperature over a period of time, temperature deviation;

[0036] Historical load characteristics: based on historical net load Select the load of the same period a few days before the forecast time as the load value of the same period; calculate the standard deviation of the load of the same period a few days before the forecast time.

[0037] 2) XGBoost point prediction

[0038] The XGBoost model is used for training with the aforementioned feature parameters to output a load baseline prediction. Set the number of decision trees, maximum tree depth, and learning rate, and ensure the accuracy of point prediction through cross-validation;

[0039] 3) Timing error filtering intervalization

[0040] Calculate the actual electricity load of users Compared with the load baseline forecast error : Errors were processed using a 3-point moving average filter. : In the formula, for Time error, for Time error, This represents the filtered error.

[0041] Then, based on the load fluctuation level of the current seasonal zone, select the upper and lower quantiles of the filtered error. , Finally, the load forecast range is obtained, and its lower limit is denoted as . The upper limit is denoted as : ;

[0042] 3.3 Construction of Photovoltaic Output Range Prediction Model

[0043] 1) Feature Engineering

[0044] Based on historical net load data Irradiance Ambient temperature From the raw time-series data, the following types of characteristic parameters strongly correlated with photovoltaic power output were extracted to ensure they match physical laws:

[0045] Meteorological characteristics: solar irradiance Ambient temperature and rate of change of radiation ;

[0046] Historical output characteristics: Based on the theoretical estimate of photovoltaic output Based on this, we obtain the average photovoltaic power output of the same period in the days preceding the predicted time and the power output value of the hours preceding the predicted time.

[0047] System characteristics: installed capacity S, temperature-corrected conversion efficiency ;

[0048] 2) LGBM Lightweight Point Prediction

[0049] The LGBM model is used to determine the benchmark value of photovoltaic power output. Point prediction; set the number of leaf nodes, maximum tree depth, and learning rate; model training aims to minimize the mean squared error (MSE);

[0050] 3) Dynamic error pool guided by physical rules + adaptive quantile

[0051] Establish three quantifiable physical rules to filter for reasonable errors:

[0052] Rule 1: Irradiation-output positive correlation rule

[0053] If the current light irradiance Not lower than the irradiance of the previous moment Therefore, the photovoltaic output should not drop sharply, satisfying the following condition: Corresponding error constraints: In the formula, for Theoretical estimate of time; It is the compliance prediction error after being filtered by physical rules;

[0054] Rule 2: High Temperature Efficiency Reduction Constraint Rule

[0055] When the ambient temperature exceeds 35℃, the theoretical estimated value of photovoltaic output will decrease due to the degradation of photovoltaic system efficiency. It should not exceed the photovoltaic power output benchmark value :like ,but ;

[0056] Rule 3: Output Extreme Boundary Rule

[0057] Photovoltaic power output must meet the physical system boundary conditions: Corresponding error constraints: ; It is the compliance prediction error after being filtered by physical rules;

[0058] Calculate the original error Construct recent historical data, and use the recent historical data to construct and dynamically update an error pool that contains only compliant raw errors;

[0059] Based on the irradiance level for the predicted period, the high and low quantiles selected from the error pool are dynamically adjusted. , The final photovoltaic output prediction range is generated, and its lower limit is denoted as . The upper limit is denoted as : ,

[0060] If the lower limit of the interval <0, adjust to 0; if the upper limit of the interval > Adjustments are made according to the rated values ​​to ensure full compliance with photovoltaic operating principles.

[0061] The above-mentioned method for predicting net load intervals for users with distributed photovoltaic power generation based on physical constraints involves the following specific process for generating the net load prediction interval:

[0062] Step 4: Net Load Interval Fusion Method under Physical Constraints

[0063] 4.1 Quantitative Calculation of the Correlation among the Three

[0064] The Copula function is used to quantify the dynamic correlation between the load forecast interval, the photovoltaic output forecast interval, and the net load historical interval. The joint probability distribution correlation degree of the three is calculated. The nonlinear correlation relationship of the three is fitted by the Copula function, and the correlation quantification results of winter load and net load with a correlation degree ≥0.8 and summer photovoltaic output and net load with a correlation degree ≤-0.7 are output, providing a basis for weight allocation.

[0065] 4.2 Dynamic Weight Allocation

[0066] Based on the correlation quantification results, a weight allocation strategy that dynamically adjusts with lighting scenarios and seasons is constructed, and a preliminary estimate of the net load range is obtained through weighted calculation: ,in, and The calculated lower and upper limits of the net load range are estimated. and These represent the dynamic weights of the load forecast interval and the photovoltaic output forecast interval in the integration process, respectively. and These are the lower and upper limits of the load forecast range; and This represents the lower and upper limits of the photovoltaic power output prediction range.

[0067] The aforementioned method for predicting user net load intervals with distributed photovoltaic power based on physical constraints further includes, in step four:

[0068] 4.3 Seasonal Characteristics Correction

[0069] To improve the seasonal adaptability of the fusion results, a seasonal correction factor is introduced. The seasonal correction factor The systematic deviation was calculated based on historical data from each quarter:

[0070] First calculate the... Quarterly average absolute deviation rate This is used to quantify the systematic bias in historical forecasts for this season: In the formula, For the first The number of historical data points included in a quarter; and Representing the first The lower and upper limits of the load forecast interval corresponding to each historical data point; and Representing the first The lower and upper limits of the photovoltaic power output prediction range corresponding to each historical data point; Then it is the first The actual net load value corresponding to each historical data point;

[0071] Based on mean absolute deviation , define the first Quarterly seasonal adjustment factor for: , The shrinkage coefficient is used to adjust the initial fusion results, resulting in the seasonally corrected interval: , and The lower and upper limits of the revised net load range are estimated.

[0072] The aforementioned method for predicting user net load intervals with distributed photovoltaic power based on physical constraints further includes, in step four:

[0073] 4.4 Physical Constraint Verification

[0074] Based on the physical principle that net load = load - photovoltaic output, the seasonally corrected range is finally verified and subject to hard constraints to ensure that it meets the theoretical extreme value:

[0075] Upper limit constraint: If Then take ;

[0076] Lower bound constraint: If Then take ;

[0077] 4.5 Adaptive Optimization Mechanism

[0078] By employing a rolling time window mechanism to periodically update parameters, this method possesses long-term adaptive capabilities.

[0079] Compared with the prior art, the present invention has the following significant advantages:

[0080] First, it has high decoupling accuracy and is non-invasive and low-cost: Through the "physical mechanism + data-driven" model and residual distribution analysis, it achieves accurate decoupling based solely on grid connection point data, without the need for additional metering devices, making it low-cost and easy to implement.

[0081] Second, the interval forecasts are realistic and have strong coverage: Different interval forecast models were designed for the different characteristics of load and photovoltaic output, abandoning the assumption of a single distribution. The generated intervals can more realistically reflect uncertainty, especially with better coverage of extreme fluctuation events.

[0082] Third, the fusion process is scientific and has strong physical interpretability: It innovatively introduces Copula correlation analysis for dynamic weight allocation and applies strict physical constraints to make the fusion results not only data-driven but also guided by physical laws, ensuring the rationality and reliability of the intervals.

[0083] Fourth, Adaptability and Refinement: Through seasonal partitioning, scene recognition, and rolling optimization mechanisms, the model can finely adapt to changes in characteristics at different time scales, possessing strong adaptive capabilities and ensuring the long-term effectiveness of the method in changing environments. Attached Figure Description

[0084] Figure 1 This is an overall flowchart of the net load interval prediction method of the present invention.

[0085] Figure 2 Figure showing the modeling and decoupling verification results for the theoretical value of photovoltaic power output.

[0086] Figure 3 This is a model architecture diagram for inter-item interval prediction. Detailed Implementation

[0087] A method for predicting the net load interval of users with distributed photovoltaic power generation based on physical constraints includes the following steps:

[0088] Step 1: Data Acquisition and Preprocessing

[0089] Collect historical net load of user and grid connection points Irradiance Ambient temperature And the basic parameters of the photovoltaic system, including installed capacity. Rated conversion efficiency Temperature coefficient .

[0090] Data preprocessing methods (such as...) (Criteria) Remove outliers; for short-term missing data, use interpolation methods (such as linear interpolation) to complete the data and ensure data continuity.

[0091] Step 2: Verification of the accuracy of load and photovoltaic output separation and decoupling.

[0092] This step aims to solve the problem of aliasing modeling of load and photovoltaic output by combining physical mechanisms with data-driven approaches to achieve non-intrusive and precise decoupling between the two.

[0093] 2.1 Modeling of Theoretical Photovoltaic Output

[0094] Based on the IV characteristic curve of the photovoltaic system, the irradiance is fitted using the random forest algorithm. Ambient temperature Theoretical estimates of photovoltaic output The nonlinear relationship between light, temperature and photovoltaic output is used to construct a three-dimensional mapping model.

[0095] The input features of the 3D mapping model include irradiance. Ambient temperature and temperature-corrected conversion efficiency The output is a theoretical estimate of the photovoltaic power output. The calculation formula is:

[0096] ,

[0097] in, The calculation formula is: ,

[0098] In the formula, For rated conversion efficiency, For temperature coefficient, For ambient temperature, This is a reference temperature (usually 25°C).

[0099] The theoretical estimate of photovoltaic power output was fitted using the random forest algorithm. Compared with historical actual photovoltaic output To mitigate bias, optimize model parameters (number of decision trees, maximum tree depth) to reduce the theoretical estimate of photovoltaic output. Compared with historical actual photovoltaic output To minimize deviations, ensure that the model fit is R² ≥ 0.9.

[0100] 2.2 Reverse splitting of load and photovoltaic output

[0101] Based on historical net load Theoretical estimates of photovoltaic power output Based on this, a dynamic correction term is introduced. After eliminating photovoltaic fluctuations, the actual electricity load of users is estimated by reverse derivation. Residual analysis was performed within a sliding time window (15 minutes). The actual electricity load of users was estimated using a formula. :

[0102] ,

[0103] Among them, dynamic correction items Based on deviation ( The probability distribution of ) is dynamically determined.

[0104] 2.3 Decoupling Quality Assessment and Model Iteration

[0105] To evaluate the rationality of the non-invasive decoupling results, a dual evaluation mechanism is constructed.

[0106] 1) Physical rationality test

[0107] Decoupled user actual power load It satisfies basic physical laws: its value is non-negative for the vast majority of the time period; its instantaneous rate of change Within a reasonable range for the application electricity scenario, avoid physical power surges; the load curve should exhibit a daily periodicity consistent with user behavior patterns, for example, the nighttime load should be maintained at a baseline level significantly lower than the daytime load.

[0108] 2) Residual sequence analysis

[0109] Reconstructing residuals by calculating net load A quantitative assessment of the overall consistency of decoupling is performed. Net load reconfiguration residuals. Defined as:

[0110] ,

[0111] Residual sequence mean with standard deviation It must meet the preset standards.

[0112] If the above test fails, decoupling iterative optimization is triggered.

[0113] Step 3: Build an interval prediction model

[0114] This step addresses the actual power load of the user after decoupling. Theoretical estimates of photovoltaic output Interval prediction models for load and photovoltaic output were constructed separately to address the problem that a single distribution assumption cannot accurately reflect actual fluctuation characteristics.

[0115] 3.1 Quantitative Division of Seasonal Zones

[0116] To refine the modeling of the differentiated impacts of climate and behavior on load and photovoltaic output, the year is divided into several typical seasonal zones based on temperature and electricity consumption patterns: the main season is divided into winter, summer, and spring / autumn based on daily average temperature; each main season is further subdivided into peak electricity consumption sub-zones and stable electricity consumption sub-zones based on electricity consumption intensity.

[0117] 3.2 Construction of Load Interval Prediction Model

[0118] An XGBoost point prediction + temporal error filtering interval model is constructed to balance behavioral pattern capture and uncertainty quantification, as detailed below:

[0119] 1) Feature Engineering

[0120] Based on historical net load data Irradiance Ambient temperature From the raw time-series data, multiple characteristic parameters highly correlated with electricity load are filtered to balance interpretability and predictability, avoiding redundancy.

[0121] Time features: Extract hour codes, date types (weekdays / weekends / holidays), and seasonal partition identifiers from time-series data;

[0122] Environmental characteristics: average temperature over a period of time, temperature deviation (the difference between the current temperature and the seasonal average temperature).

[0123] Historical load characteristics: based on historical net load The load of the same period one day before the forecast time is selected as the load value of the same period; the standard deviation of the load of the same period three days before the forecast time is calculated to characterize the load fluctuation level.

[0124] 2) XGBoost point prediction

[0125] A simplified XGBoost model is used, trained with the aforementioned feature parameters, to output a load baseline prediction. Set the number of decision trees, maximum tree depth, and learning rate (50, 4, 0.1), and use cross-validation to ensure point prediction accuracy (e.g., mean squared error). ).

[0126] 3) Timing error filtering intervalization

[0127] Abandoning traditional fixed statistical methods, this method quantifies load uncertainty through time-series filtering and quantile calibration to adapt to behavioral fluctuation characteristics.

[0128] Calculate the actual electricity load of users Compared with the load baseline forecast error :

[0129] ,

[0130] Error was handled using a "3-point moving average filter". :

[0131] ,

[0132] In the formula, This is the error from the previous moment. For the error at the next moment, This represents the filtered error.

[0133] Then, based on the load fluctuation level of the current seasonal zone, select the upper and lower quantiles of the filtered error. , For example, high volatility partitioning and , middle fluctuation partitioning and Low volatility partitioning and The final load range is obtained, and its lower limit is denoted as... The upper limit is denoted as :

[0134] ,

[0135] If the lower limit of the interval is less than 0, adjust it to 0 (non-negative load); if the upper limit of the interval exceeds 1.1 times the historical maximum load of the partition, adjust it to 1.1 times the historical maximum load.

[0136] 3.3 Construction of Photovoltaic Output Range Prediction Model

[0137] This step introduces physical rule constraints to construct an LGBM point prediction + physically rule-guided dynamic error pool + adaptive quantile model:

[0138] 1) Feature Engineering

[0139] Based on historical net load data Irradiance Ambient temperature From the raw time-series data, the following types of features strongly correlated with photovoltaic power output were extracted to ensure they match physical laws:

[0140] Meteorological characteristics: solar irradiance Ambient temperature Irradiation variation rate (Reflecting the fluctuations in radiation caused by cloud movement);

[0141] Historical output characteristics: Based on the theoretical estimate of photovoltaic output Based on this, the average photovoltaic power output for the same period of the previous 3 days and the power output for the first hour were obtained;

[0142] System characteristics: installed capacity S, temperature-corrected conversion efficiency .

[0143] 2) LGBM Lightweight Point Prediction

[0144] The LGBM regression model was used to determine the baseline value of photovoltaic power output. Point prediction. Set the number of leaf nodes, maximum tree depth, and learning rate (30, 3, 0.1); model training uses minimizing the mean squared error (MSE) as the objective function.

[0145] 3) Dynamic error pool guided by physical rules + adaptive quantile

[0146] Establish three quantifiable physical rules to filter for reasonable errors:

[0147] Rule 1: Irradiation-output positive correlation rule

[0148] If the current light irradiance If the power output is not lower than the previous moment, then the photovoltaic output should not drop sharply and must meet the following requirements:

[0149] ,

[0150] Corresponding error constraints: ,

[0151] In the formula, for Theoretical estimate of time; It is the compliance prediction error after being filtered by physical rules;

[0152] Rule 2: High Temperature Efficiency Reduction Constraint Rule

[0153] When the ambient temperature exceeds 35℃, the theoretical estimated value of photovoltaic output will decrease due to the degradation of photovoltaic system efficiency. It should not exceed the photovoltaic power output benchmark value :

[0154] like ,but ,

[0155] Rule 3: Output Extreme Boundary Rule

[0156] Photovoltaic power output must meet the physical system boundary conditions:

[0157] ,

[0158] Corresponding error constraints: , It is the compliance prediction error after being filtered by physical rules;

[0159] Calculate the original error Construct 15 days of recent historical data, and use the recent historical data to build and dynamically update an error pool that contains only compliant raw errors.

[0160] Based on the solar irradiance intensity level (e.g., high, medium, low) for the predicted period, the high and low quantiles selected from the error pool are dynamically adjusted. , For example, high-irradiation extraction and , medium irradiation and Low radiation and ;

[0161] The final photovoltaic output prediction range is generated, and its lower limit is denoted as . The upper limit is denoted as :

[0162] ,

[0163] If the lower limit of the interval is less than 0, adjust it to 0; if the upper limit of the interval is greater than 0, adjust it to 0. Adjustments are made according to the rated values ​​to ensure full compliance with photovoltaic operating principles.

[0164] Step 4: Net Load Interval Fusion Method under Physical Constraints

[0165] This step aims to address the problem of neglecting correlation in simple interval superposition, and to achieve the scientific integration of net load intervals by quantifying correlation and physical constraints.

[0166] 4.1 Quantitative Calculation of the Correlation among the Three

[0167] The Copula function is used to quantitatively analyze the dynamic correlation between the load forecast interval, the photovoltaic output forecast interval, and the net load historical interval. The joint probability distribution correlation degree of the three is calculated. By selecting the Gaussian Copula function to fit the nonlinear correlation of the three, the quantitative results of the correlation degree between winter load and net load ≥0.8 and the correlation degree between summer photovoltaic output and net load ≤-0.7 are output. (In winter, photovoltaic output is weak and net load is dominated by electricity load, and the two fluctuate in the same direction; in summer, photovoltaic output is significant and has an inverse relationship with net load, which is mutually exclusive.) This provides a basis for weight allocation.

[0168] 4.2 Dynamic Weight Allocation

[0169] Based on the correlation quantification results obtained in Section 4.1, a weight allocation strategy that dynamically adjusts with the lighting scenario and season is constructed (for example, in a cloudy scenario, photovoltaic output is the main source of net load uncertainty, so the photovoltaic output prediction range is given a higher weight; in a winter scenario, electricity load is the main determinant of net load level and fluctuation, so the load prediction range is given a higher weight). A preliminary estimate of the net load range is obtained through weighted calculation:

[0170] ,

[0171] in, and The calculated lower and upper limits of the net load range are estimated. and These represent the dynamic weights of the load forecast interval and the photovoltaic output forecast interval in the integration process, respectively. and These are the lower and upper limits of the load forecast range; and This represents the lower and upper limits of the photovoltaic power output prediction range.

[0172] 4.3 Seasonal Characteristics Correction

[0173] To improve the seasonal adaptability of the fusion results, a seasonal correction factor is introduced. Seasonal correction factor The systematic deviation was calculated based on historical data from each quarter:

[0174] First calculate the... Quarterly average absolute deviation rate This is used to quantify the systematic bias in historical forecasts for this season:

[0175]

[0176] In the formula, For the first The number of historical data points included in a quarter; and Representing the first The lower and upper limits of the load forecast interval corresponding to each historical data point; and Representing the first The lower and upper limits of the photovoltaic power output prediction range corresponding to each historical data point; Then it is the first The actual net load value corresponding to each historical data point.

[0177] Based on historical seasonal average absolute deviation , define the first Quarterly seasonal adjustment factor for:

[0178] ,

[0179] As the shrinkage coefficient (0 < ≤1), used to adjust the initial fusion results to obtain the seasonally corrected interval:

[0180] ,

[0181] in, and The lower and upper limits of the revised net load range are estimated;

[0182] 4.4 Physical Constraint Verification

[0183] Based on the physical principle that net load = load - photovoltaic output, the seasonally corrected range is finally verified and subject to hard constraints to ensure that it meets the theoretical extreme value:

[0184] Upper limit constraint: If Then take .

[0185] Lower bound constraint: If Then take .

[0186] This step ensures the reasonableness of the output range.

[0187] 4.5 Adaptive Optimization Mechanism

[0188] By employing a rolling time window mechanism to periodically update parameters, this method possesses long-term adaptive capabilities. It can continuously track changes in load habits and photovoltaic system performance, achieving seamless adaptation during critical periods such as seasonal transitions without manual intervention.

[0189] The present invention will be further described in detail below through specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0190] In the implementation of step one, outlier removal specifically adopts... The rule is that missing value imputation uses linear interpolation for missing data segments with a duration of no more than 2 hours.

[0191] In step two, the machine learning algorithm used for modeling the theoretical photovoltaic output is the random forest algorithm, with 100 decision trees and a maximum tree depth of 10. In the decoupling quality assessment, the residual sequence must satisfy the mean... Standard deviation .

[0192] In the implementation of step three, the seasonal zones are divided as follows: the main season is based on the daily average temperature, winter ( ),summer( ), spring and autumn ( Each main season is further subdivided into peak electricity consumption sub-segments (such as 6:00-9:00 AM and 6:00-10:00 PM in winter) and stable electricity consumption sub-segments (such as 12:00-3:00 PM in winter).

[0193] In the load interval forecasting model, point forecasting uses the XGBoost model with the following parameters: number of decision trees = 50, maximum tree depth = 4, and learning rate = 0.1. Training and testing are conducted using historical data from the past 6 months, with a minimum mean square error (MSE) of ≤0.5 kW² for the test set point forecasts. Interval forecasting is achieved through quantile calibration, dynamically selecting the error quantiles based on seasonal fluctuation levels: 2% and 98% quantiles for high fluctuation periods (winter and summer peaks), 3% and 97% quantiles for medium fluctuation periods (spring and autumn peaks), and 5% and 95% quantiles for low fluctuation periods (stable periods).

[0194] In the photovoltaic interval prediction model, the point prediction adopts the LGBM model with the following parameters: number of leaf nodes = 31, maximum tree depth = 3, and learning rate = 0.1. Historical data from the past 3 months are used for training and testing. The interval prediction is based on a dynamic error pool and an adaptive quantile method: the error pool is initialized with the compliant error of the past 15 days, updated daily, and retains the data of the most recent 30 days; the quantiles are adaptively adjusted according to the irradiance intensity, with the 3% and 97th quantiles used for high irradiance (≥800W / m²), the 2% and 98th quantiles used for medium irradiance (300-800 W / m²), and the 1% and 99th quantiles used for low irradiance (≤300W / m²).

[0195] In the implementation of step four, the dynamic weight allocation strategy is specifically as follows: weighting of photovoltaic power output prediction intervals in multi-cloud scenarios. Load forecast interval weight In a sunny day scene , Winter scenes , The adaptive optimization mechanism uses a 7-day rolling time window.

Claims

1. A method for interval forecasting of a user's net load with distributed photovoltaic based on physical law constraints, characterized in that: First, a non-intrusive decoupling of the solar-driven process is performed, splitting the net load into independent theoretical estimates of actual user electricity load and photovoltaic output. Then, considering the distinct fluctuation characteristics of actual user electricity load and photovoltaic output, suitable interval prediction models are constructed for each. The load prediction interval is obtained by focusing on user behavior patterns on the load side, while the photovoltaic output prediction interval is obtained by embedding physical constraints on the photovoltaic side. Finally, based on the load prediction interval and the photovoltaic output prediction interval, and through a dynamic fusion strategy guided by both quantitative correlation and physical laws, a net load prediction interval that conforms to both statistical data characteristics and strictly adheres to physical laws is generated.

2. The method of claim 1, wherein the method is characterized by: The specific process of non-intrusive decoupling includes: Step 1: Data Acquisition and Preprocessing Collect historical net load of user and grid connection points Irradiance Ambient temperature And the basic parameters of the photovoltaic system, including installed capacity. Rated conversion efficiency and temperature coefficient ; Step 2: Decomposing the load and photovoltaic output 2.1 Modeling of Theoretical Photovoltaic Output Based on the IV characteristic curve of the photovoltaic system, the irradiance is fitted using the random forest algorithm. Ambient temperature Theoretical estimates of photovoltaic output The nonlinear relationship between them is used to construct a three-dimensional mapping model; the input features of the three-dimensional mapping model include irradiance. Ambient temperature and temperature-corrected conversion efficiency The output is a theoretical estimate of the photovoltaic power output. , ,in, In the formula, For reference temperature; 2.2 Reverse splitting of load and photovoltaic output Based on historical net load Theoretical estimates of photovoltaic power output Based on this, a dynamic correction term is introduced. After eliminating photovoltaic fluctuations, the actual electricity load of users is estimated by reverse derivation. , .

3. The method of claim 2, wherein the method is characterized by: Step two above also includes: 2.3 Decoupling Quality Assessment and Decoupling Iteration To assess the validity of the non-invasive decoupling results, a dual evaluation mechanism is constructed. 1) Physical rationality test Decoupled user actual power load It conforms to physical laws: its value is non-negative for the vast majority of the time period; its instantaneous rate of change Within a reasonable range for the application power scenario, avoid physically impossible power surges; the load curve should exhibit a daily periodicity consistent with user behavior patterns; 2) Residual sequence analysis Reconstructing residuals by calculating net load Quantitatively assess the overall consistency of decoupling, and evaluate the net load reconstruction residual. Defined as: , the residual sequence mean with standard deviation Preset standards must be met; If the above test fails, decoupling iterative optimization is triggered.

4. The method of claim 2 or 3, wherein the method is characterized by: In step 2.1, the theoretical estimates of the photovoltaic output are fitted with a random forest algorithm deviation from the historical actual photovoltaic output The model parameters are optimized to reduce the deviation of the theoretical estimates of the photovoltaic output from the historical actual photovoltaic output The model fit goodness R2 is ensured to be ≥ 0.

9.

5. The method of claim 3, wherein the method is characterized by: The specific process of constructing an adapted interval prediction model is as follows: Step 3: Build an interval prediction model 3.1 Quantitative Division of Seasonal Zones To refine the modeling of the differentiated impacts of climate and user behavior on load and photovoltaic output, the year is divided into several typical seasonal zones based on temperature and electricity consumption patterns: the main season is divided into winter, summer, and spring / autumn based on daily average temperature; each main season is further subdivided into peak electricity consumption sub-zones and stable electricity consumption sub-zones based on electricity consumption intensity. 3.2 Construction of Load Interval Prediction Model An XGBoost point prediction + temporal error filtering interval model is constructed to balance behavioral pattern capture and uncertainty quantification, as detailed below: 1) Feature Engineering Based on historical net load data Irradiance Ambient temperature Raw time-series data, filtered for multiple characteristic parameters highly correlated with electricity load: Time features: Extract hour codes, date types, and seasonal partition identifiers from time-series data; Environmental characteristics: average temperature over a period of time, temperature deviation; Historical load characteristics: based on historical net load The load at the same period in the days before the prediction time is selected as the same period load value; the standard deviation of the load at the same period in the days before the prediction time is calculated; 2) XGBoost point prediction The XGBoost model is used for training with the aforementioned feature parameters to output a load baseline prediction. Set the number of decision trees, maximum tree depth, and learning rate, and ensure the accuracy of point prediction through cross-validation; 3) Timing error filtering intervalization Calculate the user's actual electricity load Compared with the load baseline forecast error : Errors were processed using a 3-point moving average filter. : In the formula, for Time error, for Time error, This represents the filtered error. Then, based on the load fluctuation level of the current seasonal zone, select the upper and lower quantiles of the filtered error. , Finally, the load forecast range is obtained, and its lower limit is denoted as . The upper limit is denoted as : ; 3.3 Construction of Photovoltaic Output Range Prediction Model 1) Feature Engineering Based on the collected historical net load , light irradiance , ambient temperature The original time series data, the following categories of strong correlation with photovoltaic output features are extracted to ensure that they match the physical law: Meteorological characteristics: light irradiance , ambient temperature and irradiance rate of change ; history contribute Features: In the theoretical estimate of photovoltaic output Based on this, we obtain the average photovoltaic power output of the same period in the days preceding the predicted time and the power output value of the hours preceding the predicted time. System characteristics: installed capacity S, temperature-corrected conversion efficiency ; 2) LGBM Lightweight Point Prediction Point prediction of photovoltaic power output reference values using LGBM model ​ Set the number of leaf nodes, maximum tree depth, and learning rate, and train the model with the goal of minimizing the mean squared error (MSE). 3) Dynamic error pool guided by physical rules + adaptive quantile Establish three quantifiable physical rules to filter for reasonable errors: Rule 1: Irradiation-output positive correlation rule If the current light irradiance Not lower than the irradiance of the previous moment Therefore, the photovoltaic output should not drop sharply, satisfying the following condition: Corresponding error constraints: In the formula, for Theoretical estimate of time; It is the compliance prediction error after being filtered by physical rules; Rule 2: High Temperature Efficiency Reduction Constraint Rule When the ambient temperature exceeds 35℃, the theoretical estimated value of photovoltaic output will decrease due to the degradation of photovoltaic system efficiency. It should not exceed the photovoltaic power output benchmark value :like ,but ; Rule 3: Output Extreme Boundary Rule The photovoltaic power output needs to meet the physical system boundary conditions: , the corresponding error constraint: ; is the compliant prediction error filtered by physical rules; Calculate the original error Construct recent historical data, and use the recent historical data to construct and dynamically update an error pool that contains only compliant raw errors; Based on the irradiance level for the predicted period, the high and low quantiles selected from the error pool are dynamically adjusted. , The final photovoltaic output prediction range is generated, with its lower limit denoted as . The upper limit is denoted as : , If interval lower limit < 0, adjust to 0; if interval upper limit > , correct by rated value, ensure complete compliance with photovoltaic operation rules.

6. The method of claim 5, wherein the method is characterized by: The specific process for generating the net load forecast interval is as follows: Step 4: Net Load Interval Fusion Method under Physical Constraints 4.1 Quantitative Calculation of the Correlation among the Three The Copula function is used to quantify the dynamic correlation between the load forecast interval, the photovoltaic output forecast interval, and the net load historical interval. The joint probability distribution correlation degree of the three is calculated. The nonlinear correlation relationship of the three is fitted by the Copula function, and the correlation quantification results of winter load and net load with a correlation degree ≥0.8 and summer photovoltaic output and net load with a correlation degree ≤-0.7 are output, providing a basis for weight allocation. 4.2 Dynamic Weight Allocation Based on the correlation quantification results, a weight allocation strategy that dynamically adjusts with lighting scenarios and seasons is constructed, and a preliminary estimate of the net load range is obtained through weighted calculation: ,in, and The calculated lower and upper limits of the net load range are estimated. and These represent the dynamic weights of the load forecast interval and the photovoltaic output forecast interval in the integration process, respectively. and These are the lower and upper limits of the load forecast range; and This represents the lower and upper limits of the photovoltaic power output prediction range.

7. The method for predicting user net load intervals with distributed photovoltaic power based on physical law constraints according to claim 6, characterized in that: Step four also includes: 4.3 Seasonal Characteristics Correction To improve the season adaptability of the fusion result, a season correction factor is introduced The season correction factor is calculated based on the systematic bias of historical data of each season. First calculate the... Quarterly average absolute deviation rate This is used to quantify the systematic bias in historical forecasts for this season: In the formula, For the first The number of historical data points included in a quarter; and Representing the first The lower and upper limits of the load forecast interval corresponding to each historical data point; and Representing the first The lower and upper limits of the photovoltaic power output prediction range corresponding to each historical data point; Then it is the first The actual net load value corresponding to each historical data point; Based on mean absolute deviation , define the first Quarterly seasonal adjustment factor for: , The shrinkage coefficient is used to adjust the initial fusion results, resulting in the seasonally corrected interval: , and The lower and upper limits of the revised net load range are estimated.

8. The method of claim 7, wherein the method is characterized by: Step four also includes: 4.4 Physical Constraint Verification Based on the physical principle that net load = load - photovoltaic output, the seasonally corrected range is ultimately verified and subject to hard constraints to ensure that it meets the theoretical extreme value: Upper limit constraint: If Then take ; Lower bound constraint: if then take ; 4.5 Adaptive Optimization Mechanism By employing a rolling time window mechanism to periodically update parameters, this method possesses long-term adaptive capabilities.

Images