Data drift detection method and system for photovoltaic power prediction model

Abstract

The invention discloses a data drift detection method and system for a photovoltaic power prediction model. The method comprises the following steps: acquiring daily actual power data, forecast irradiation and actual irradiation of a photovoltaic power generation system; calculating an IOU value of the daily actual power data and preset standard power data as a first IOU value; calculating the IOU value of the forecast irradiation and the actual irradiation of each date, and taking the IOU value as a second IOU value; when the first IOU value and the second IOU value are both larger than the corresponding preset threshold values, the day is used as a detection day, a unified normalization factor is established based on the station grid-connected capacity and the actual irradiation maximum value, and the IOU value of the actual power and forecast irradiation corresponding to the detection day after normalization is obtained and is used as a third IOU value; and when the change amplitude of the third IOU value between two adjacent detection days is greater than the preset standard amplitude, judging that the photovoltaic power prediction model has data drift. The stability and reliability of the photovoltaic power prediction model can be improved.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of photovoltaic power generation, and in particular to a data drift detection method and system for a photovoltaic power prediction model. Background Art

[0002] With the increasing application of photovoltaic power generation systems, more and more photovoltaic power sources are connected to the distribution network, which brings huge challenges to the planning, operation, control and other aspects of the power system. Since the amount of solar radiation is closely related to meteorological conditions, the output power of photovoltaic power generation systems is inherently random and volatile. In the case that the mismatch between power storage facilities and new energy grid-connected power is difficult to change in the short term, the access of large-scale photovoltaic power generation systems to the power grid will have a great impact on the safe and stable operation of the power system. This is also a key technical problem that needs to be solved for the large-scale access of photovoltaic power generation to the power grid. Countries around the world have successively carried out technical research on photovoltaic power generation power prediction, which is of great significance to the stable operation of the power system, and helps the power system dispatching department to coordinate the power generation planning of conventional energy and photovoltaic power generation, and reasonably arrange the operation mode of the power grid.

[0003] At present, most of the models used to predict photovoltaic power generation are machine learning models and deep learning models. The existing photovoltaic power generation prediction model constructs a mapping relationship between forecast weather, historical meteorological data and historical power generation data to power generation of power stations, and uses this mapping relationship to predict the power generation of photovoltaic stations in future time periods. Due to changes in data distribution, the mapping relationship between input and output will change over time. This change may cause the learning model to not match the current data distribution, thereby affecting the prediction performance of the model. We call this phenomenon data drift. Scholars have proposed a variety of methods for detecting whether data drift occurs, such as drift detection method (DDM), early drift detection method (EDDM), detection method using statistical testing (STEPD), adaptive window-based drift detection method (ADDM), fast Hoeffding drift detection method (FHDDM), etc. This type of method mainly determines whether data drift occurs by detecting indicators that describe the performance of the model (such as accuracy) or changes in input data.

[0004] In real business scenarios, the forecast irradiance will have obvious seasonal changes, that is, it is lower in winter and higher in summer. However, the actual power of photovoltaic power stations does not change strongly with seasonality, so there will be two dates with very close actual power, and the curve of the predicted irradiance intensity may have significant deviations, which will then lead to corresponding differences in the power values ​​predicted by the model.

[0005] Some studies have proposed some methods to preprocess the forecast irradiation data to reduce the seasonal variation of forecast irradiation, but they have not been able to completely eliminate it. The existing method is to monitor the performance of the model and update the model based on the new data if significant changes are found in the performance of the model.

[0006] Existing methods mainly determine whether data drift occurs by detecting the performance indicators of the model. Considering that there are two sources of error affecting the model, one is the forecast meteorological source deviation and the other is the model itself. Existing methods only detect model performance and cannot well separate the impact of the model itself from the meteorological source deviation. Summary of the invention

[0007] In view of the technical problems existing in the prior art, the present invention provides a data drift detection method and system for a photovoltaic power prediction model, which can improve the stability and reliability of the photovoltaic power prediction model.

[0008] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0009] A method for detecting data drift of a photovoltaic power prediction model comprises the following steps:

[0010] Obtain actual power data, forecast irradiation and actual irradiation of photovoltaic power generation system every day;

[0011] Calculate the IOU value of the actual power data of each day and the preset standard power data as the first IOU value; calculate the IOU value of the forecast irradiation and the actual irradiation of each date as the second IOU value;

[0012] Compare the first IOU value and the second IOU value of each day with the corresponding preset threshold value; when the first IOU value and the second IOU value are both greater than the corresponding preset threshold value, take this day as the detection day, and establish a unified normalization factor based on the grid-connected capacity of the site and the actual irradiation maximum value, and obtain the IOU value of the actual power and the predicted irradiation corresponding to the normalization of the detection day as the third IOU value;

[0013] Calculate the change range of the third IOU value between two adjacent detection days; compare the change range of the third IOU value between two adjacent detection days with the preset standard range; when the change range of the third IOU value between two adjacent detection days is greater than the preset standard range, it is determined that data drift occurs in the photovoltaic power prediction model.

[0014] Preferably, the preset standard power data is data corresponding to a seed date.

[0015] Preferably, the seed date selection process is:

[0016] Calculate the total power generation of the photovoltaic power generation system every day, the IOU value of the predicted irradiation and the actual irradiation, and the IOU value of the predicted power and the actual power;

[0017] Determine whether the total power generation of each day exceeds the set threshold; determine whether the IOU value of the forecast irradiation and the actual irradiation of each day exceeds the set threshold; determine whether the IOU value of the forecast power and the actual power of each day exceeds the set threshold;

[0018] When the total power generation of each day exceeds the set threshold, and the IOU value of the forecast irradiation and the actual irradiation of each day exceeds the set threshold, and the IOU value of the forecast power and the actual power of each day exceeds the set threshold, the data of this day is recorded and summarized to obtain all the summary data; the summary data includes the date and the total power generation corresponding to each date;

[0019] The aggregated data are sorted by the total power generation value, and the date corresponding to the median total power generation is used as the seed date.

[0020] Preferably, the formula for calculating the variation range of the third IOU value between two adjacent detection days is:

[0021]

[0022] Where IOU i+1 , IOU i The third IOU value of two adjacent detection days.

[0023] Preferably, the process of selecting the preset standard amplitude is:

[0024] Input the actual power, forecast power, seed date and current processing date for multiple days in history;

[0025] Calculate the IOU value of the actual power corresponding to the current date and the actual power corresponding to the seed date, and determine whether the IOU value exceeds the threshold; if the IOU value exceeds the threshold, continue to calculate the IOU value of the forecast irradiation and the actual irradiation for the day, and determine whether the IOU value exceeds the threshold; if the IOU value exceeds the threshold, calculate the normalized IOU value of the forecast irradiation and the actual power for the day, and store it;

[0026] The above stored data is sorted by date, and the variation range of the IOU values ​​between two dates is calculated, and the preset standard range is obtained according to the variation range of each IOU value.

[0027] Preferably, the IOU value is the ratio of the intersection area to the union area of ​​two curves.

[0028] The present invention also discloses a computer program product, comprising a computer program, wherein the computer program executes the steps of the method described above when executed by a processor.

[0029] The present invention further discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described above are executed.

[0030] The present invention also discloses a data drift detection system for a photovoltaic power prediction model, comprising a memory and a processor connected to each other, wherein a computer program is stored in the memory, and when the computer program is run by the processor, the steps of the method described above are executed.

[0031] Compared with the prior art, the advantages of the present invention are:

[0032] The detection method of the present invention takes into account the accuracy of actual power prediction and the accuracy of irradiation prediction at the same time. By calculating the IOU value of actual power and preset standard power and the IOU value of predicted irradiation and actual irradiation, the stability of the model is evaluated from two different perspectives. By comparing with the preset threshold, it can be determined which days need further inspection; only when both IOU values ​​exceed the threshold, it is considered to be a detection day, reducing unnecessary inspections and improving efficiency. By comparing the change amplitude of the third IOU value between two adjacent detection days, it can be determined whether the model has data drift. If the change amplitude exceeds the preset standard amplitude, it can be considered that the model has data drift. This method can issue an early warning through the change of the third IOU value before the model performance decreases significantly, so as to adjust or retrain the model in time to maintain the accuracy of the prediction. Since a single IOU value is used as an evaluation indicator, the detection process is simplified, making the operation more intuitive and easy to implement. In general, the present invention can improve the stability and reliability of the photovoltaic power prediction model, reduce the prediction error caused by data drift, extend the service life of the model, and reduce maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The flowchart of the seed date selection method in the present invention is an embodiment.

[0034] Figure 2 Schematic diagram of the IOU value calculation method in the present invention; (a) is a schematic diagram of the intersection area; (b) is a schematic diagram of the union area.

[0035] Figure 3 The flowchart of the specific method for determining the preset standard amplitude in the present invention is shown in an embodiment.

[0036] Figure 4 The flowchart of the data drift detection method in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0038] The data drift detection method of the photovoltaic power prediction model provided by the embodiment of the present invention determines whether data drift occurs by detecting whether the performance of the model changes under a specific meteorological scenario.

[0039] Since the deviation of the model is determined by the deviation of the meteorological source and the performance of the model itself, the performance change of a single detection model cannot well determine whether data drift has occurred. For example, cloudy days are difficult to predict meteorological scenes, so the deviation between the forecast weather and the actual weather in cloudy scenes will be greater than the deviation between the forecast weather and the actual weather in sunny scenes. Correspondingly, the performance of the model in cloudy scenes will also be worse than that in sunny scenes. Therefore, to detect whether data drift has occurred, a fixed weather scene is required.

[0040] In the present invention, only the performance of the model in sunny scenes is tested, and then it is determined whether data drift occurs.

[0041] Before explaining how to select data for sunny scenes, we first introduce the concept of seed date, which is defined as the day when the power generation curve is representative.

[0042] The present invention only tests the performance of the model in sunny scenes, so it is necessary to first filter out historical data with sunny weather as the actual weather. However, sunny scenes are a relatively broad concept. In order to make the filtered sunny scenes closer, a representative date needs to be selected, and then other dates with similar actual weather and forecast weather to the date are selected from the data to form data for testing the performance of the model.

[0043] like Figure 1 As shown in the figure, the specific process of seed date selection is as follows:

[0044] 1. First, in the given historical data, calculate the total power generation of each day, the IOU (Intersection over Union) value of the predicted irradiation and the actual irradiation, and the IOU value of the predicted power and the actual power;

[0045] 2. Determine whether the total power generation of each day exceeds the set threshold, where the set threshold is determined based on the installed capacity of the site and weather conditions. In a sunny day scenario, the total power generation of a single day is higher than the total power generation of a cloudy or overcast day scenario, so a higher total power generation threshold is used to filter sunny days;

[0046] Determine whether the IOU value of the forecast irradiation and the actual irradiation of each day exceeds the set threshold; if the IOU value is very low, it means that the weather forecast is inaccurate and the actual power curve of the day cannot be used as a representative curve;

[0047] Determine whether the IOU value of the forecast power and the actual power of each day exceeds the set threshold; if the IOU value is very low, it means that the model forecast is inaccurate and cannot be used as a representative power curve;

[0048] The data that meets the above conditions (exceeds the corresponding set threshold) are summarized, and the summarized data includes the date and the total power generation corresponding to each date;

[0049] 3. Sort the summarized data by total power generation value, and the date corresponding to the median total power generation is the seed date.

[0050] The reason for selecting the date corresponding to the median total power generation is that the total power generation corresponding to this date will not deviate too much from the total power generation corresponding to the sunny scene on the subsequent test date. That is, it is ensured that there will be a date close to the seed date in the subsequent test date.

[0051] based on Figure 1 The seed dates selected by the flowchart shown are characterized by high total power generation, close weather forecast irradiation to actual irradiation values, and close model predicted power to actual power.

[0052] The IOU value of the two curves in a single day involved in the above process is defined as the ratio of the intersection area to the union area of ​​the two curves. The photovoltaic power curve and the irradiation curve are both 0 at night. Therefore, the power curve or irradiation curve of a single day and the time axis will form a bell-shaped surface.

[0053] like Figure 2 As shown, Figure 2 The shaded area in (a) (the position indicated by the dotted line) is the intersection area of ​​the two power curves. Figure 2 The shaded area in (b) (the position indicated by the dotted line) is the union area of ​​the two power curves. If the two power curves are closer in shape and amplitude, the IOU value is closer to 1.

[0054] like Figure 4 As shown, based on the seed date obtained by the above method, the data drift detection method of the photovoltaic power prediction model provided by the embodiment of the present invention includes the steps of:

[0055] 1. Since the meteorological scene needs to be fixed, the present invention selects a day that is close to the real power data of the seed date as the detection day. To determine whether a date can be a detection day, first calculate whether the IOU value (i.e., the first IOU value) of the real power of the day and the real power of the seed date reaches the set threshold. Only when the real power of the day is less different from the real power of the seed date, the IOU values ​​of the two may exceed the set threshold, so that the real meteorological scene of the selected detection day will be close to the meteorological scene of the seed day;

[0056] 2. At the same time, the IOU value of the forecast irradiance and the actual irradiance of the day (i.e. the second IOU value) must also exceed the set threshold. This step is to eliminate the dates with inaccurate weather forecasts;

[0057] 3. If a day meets the thresholds set in both step 1 and step 2, then that day is used as the detection day;

[0058] 4. Since the selection and testing of the test day are carried out in a rolling manner, that is, a judgment will be made every day on the previous day. Therefore, the data of multiple test days that meet the conditions will be collected, and the collected data includes the date, the corresponding actual power and the IOU value of the forecast irradiation (i.e. the third IOU value);

[0059] 5. Based on the data obtained in step 4, calculate the change in the forecast power IOU value between two consecutive detection days. If the change amplitude of the third IOU value changes significantly, for example, exceeds the preset standard amplitude (such as 20%, which can be selected between 10%-30%, depending on the actual situation), it can be determined that data drift has occurred.

[0060] The reason why the IOU value changes between the two detection days is that the two real meteorological scenes are very close, but the mapping relationship between the forecast meteorology and the real meteorological scene changes, resulting in changes in the power predicted by the same model. Therefore, there will be a difference in the IOU values ​​of the two days.

[0061] like Figure 3 As shown, specifically, the specific process of obtaining the preset standard amplitude is:

[0062] 1. Input the actual power, forecast power, seed date and current processing date of multiple days in history;

[0063] 2. Determine whether the current processing date is still within the time range covered by the input data;

[0064] 3. Calculate the IOU value of the actual power corresponding to the current date and the actual power corresponding to the seed date, and determine whether the IOU value exceeds the threshold; that is, filter whether the day is sunny; if the IOU value exceeds the threshold, execute the next step;

[0065] 4. Continue to calculate the IOU value of the forecast irradiation and the actual irradiation for the day, and determine whether the IOU value exceeds the threshold; if the IOU value exceeds the threshold, execute the next step; if the IOU value is very low, it means that the weather forecast is inaccurate and the data of the day cannot be used to calculate the preset standard range.

[0066] 5. Calculate the IOU value of the forecast irradiation and actual power for the day after normalization and store it. Since the dimensions of irradiation and power are inconsistent, they are first normalized to make the dimensions uniform, and then the IOU value of irradiation and power is calculated.

[0067] 6. Sort the stored data by date, and calculate the change range of the IOU value between two dates, and obtain the preset standard range according to the change range of each IOU value. For example, the preset standard range can be determined based on the change range obtained from the actual case, and specifically, the change range of each IOU value can be averaged to obtain the preset standard range.

[0068] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0069] Assume that model A (machine learning model or deep learning model) has been trained based on the history before January 1, 2024, and the power forecast on and after January 1, 2024 is provided by model A. When model A runs until March 1, 2024, it is necessary to determine whether data drift occurs after March 1, 2024. There is an assumption here, that is, it is assumed that model A is running stably from January 1, 2024 to March 1, 2024, and no data drift occurs. Implement this method according to the following process:

[0070] 1. First, a seed date is selected between January 1, 2024 and March 1, 2024. In this case, the seed date is January 15, 2024;

[0071] 2. On March 1, 2024 and every subsequent day, determine whether the day meets the conditions of the test day. If so, mark the day as the test day, and then calculate the IOU value of the actual power of the day and the predicted weather (specifically irradiation) of model A for the day, and record the day and the IOU value (because the dimensions of irradiation and power are inconsistent, they are first normalized to unify the dimensions, and then the IOU values ​​of irradiation and power are calculated);

[0072] 3. If the cumulative number of detection days in step 2 exceeds 2 days, the change range of the IOU value of the latest detection day and the IOU value of the previous detection day is calculated at the same time. The specific calculation formula is:

[0073]

[0074] 4. The change in the IOU value calculated in step 3 is also recorded, so the IOU value on the test day and the change in the IOU value between the test day and the previous test day are recorded at the same time;

[0075] 5. If the change in the IOU value shows a significant change, assuming it exceeds 20% (the specific threshold depends on the actual situation), it can be judged that data drift has occurred.

[0076] The IOU value-based method of detecting data drift in the present invention is mainly aimed at the scene where the weather forecast irradiance changes with the seasons. When the real weather scene is close, the IOU value change amplitude of the two detection days exceeds the threshold, indicating that there is a significant difference in the predicted power of model A for the two detection days, and the reason for this significant difference is that the mapping relationship between the weather forecast and the real power has changed, that is, data drift.

[0077] The present invention can monitor the performance of the photovoltaic power prediction model in real time by calculating the IOU value (intersection over union ratio) of the actual power and the predicted power every day; the IOU value is an indicator that measures the degree of overlap between the prediction frame and the real frame, and is used here to evaluate the accuracy of the model prediction.

[0078] The detection method of the present invention takes into account the accuracy of actual power prediction and the accuracy of irradiation prediction at the same time. By calculating the IOU value of actual power and preset standard power and the IOU value of predicted irradiation and actual irradiation, the stability of the model is evaluated from two different angles. By comparing with the preset threshold, it can be determined which days need further inspection; only when both IOU values ​​exceed the threshold, it is considered to be a detection day, which reduces unnecessary inspections and improves efficiency. By comparing the change range of the third IOU value between two adjacent detection days, it can be determined whether the model has data drift. If the change range exceeds the preset standard range, it can be considered that the model has data drift. This method can issue an early warning through the change of the third IOU value before the model performance drops significantly, so as to adjust or retrain the model in time to maintain the accuracy of the prediction. By using a single IOU value as an evaluation indicator, the detection process is simplified, making the operation more intuitive and easy to implement.

[0079] In general, the present invention can improve the stability and reliability of the photovoltaic power prediction model, reduce the prediction error caused by data drift, extend the service life of the model, and reduce maintenance costs.

[0080] The present invention also discloses a computer program product, including a computer program, which executes the steps of the above method when executed by a processor. The present invention further discloses a computer-readable storage medium, on which a computer program is stored, which executes the steps of the above method when executed by a processor. The present invention also discloses a data drift detection system for a photovoltaic power prediction model, including an interconnected memory and a processor, on which a computer program is stored, which executes the steps of the above method when executed by a processor. The product, medium and system of the present invention correspond to the above detection method and also have the advantages described in the above method.

[0081] The present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned method embodiment when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable storage media include: any entity or device that can carry computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. The memory is used to store computer programs and / or modules, and the processor implements various functions by running or executing computer programs and / or modules stored in the memory, and calling data stored in the memory. The memory may include high-speed random access memory and may also include non-volatile memory, such as a hard disk, an internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0082] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A data drift detection method for a photovoltaic power prediction model, characterized in that: Includes steps: Obtain actual power data, forecast irradiation and actual irradiation of photovoltaic power generation system every day; Calculate the IOU value of the actual power data of each day and the preset standard power data as the first IOU value; calculate the IOU value of the forecast irradiation and the actual irradiation of each date as the second IOU value; Compare the first IOU value and the second IOU value of each day with the corresponding preset threshold value; when the first IOU value and the second IOU value are both greater than the corresponding preset threshold value, take this day as the detection day, and establish a unified normalization factor based on the grid-connected capacity of the site and the actual irradiation maximum value, and obtain the IOU value of the actual power and the predicted irradiation corresponding to the normalization of the detection day as the third IOU value; Calculate the change range of the third IOU value between two adjacent detection days; compare the change range of the third IOU value between two adjacent detection days with the preset standard range; when the change range of the third IOU value between two adjacent detection days is greater than the preset standard range, it is determined that data drift occurs in the photovoltaic power prediction model.

2. The data drift detection method of the photovoltaic power prediction model according to claim 1, characterized in that: The preset standard power data is data corresponding to the seed date.

3. The data drift detection method of the photovoltaic power prediction model according to claim 2, characterized in that: The seed date selection process is as follows: Calculate the total power generation of the photovoltaic power generation system every day, the IOU value of the predicted irradiation and the actual irradiation, and the IOU value of the predicted power and the actual power; Determine whether the total power generation of each day exceeds the set threshold; determine whether the IOU value of the forecast irradiation and the actual irradiation of each day exceeds the set threshold; determine whether the IOU value of the forecast power and the actual power of each day exceeds the set threshold; When the total power generation of each day exceeds the set threshold, and the IOU value of the forecast irradiation and the actual irradiation of each day exceeds the set threshold, and the IOU value of the forecast power and the actual power of each day exceeds the set threshold, the data of this day is recorded and summarized to obtain all the summary data; the summary data includes the date and the total power generation corresponding to each date; The aggregated data are sorted by the total power generation value, and the date corresponding to the median total power generation is used as the seed date.

4. The data drift detection method of the photovoltaic power prediction model according to claim 1, 2 or 3, characterized in that: The formula for calculating the change in the third IOU value between two adjacent detection days is: Where IOU i+1 , IOU i is the third IOU value of two adjacent detection days.

5. The data drift detection method of the photovoltaic power prediction model according to claim 2 or 3, characterized in that: The process of selecting the preset standard amplitude is as follows: Input the actual power, forecast power, seed date and current date of multiple historical days; Calculate the IOU value of the actual power corresponding to the current date and the actual power corresponding to the seed date, and determine whether the IOU value exceeds the threshold; if the IOU value exceeds the threshold, continue to calculate the IOU value of the forecast irradiation and the actual irradiation for the day, and determine whether the IOU value exceeds the threshold; if the IOU value exceeds the threshold, calculate the normalized IOU value of the forecast irradiation and the actual power for the day, and store it; The above stored data is sorted by date, and the variation range of the IOU values ​​between two dates is calculated, and the preset standard range is obtained according to the variation range of each IOU value.

6. The data drift detection method of the photovoltaic power prediction model according to claim 1, 2 or 3, characterized in that: The IOU value is the ratio of the intersection area to the union area of ​​two curves.

7. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are performed.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 6.

9. A data drift detection system for a photovoltaic power prediction model, comprising a memory and a processor connected to each other, wherein a computer program is stored in the memory, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 6.

Images