Photovoltaic power station cluster power prediction method and system based on multi-task learning

Abstract

The present invention relates to the field of photovoltaic power prediction technology, specifically to a photovoltaic power station cluster power prediction method and system based on multi-task learning, comprising the following steps: obtaining the three-dimensional position and shading characteristics of the power station, constructing the solar incidence vector, identifying the array shading relationship and spatial illumination distribution, extracting the wind speed disturbance characteristics, deducing the propagation path and response timing, and performing multi-task prediction to dynamically adjust the output. In the present invention, a three-dimensional spatial shading relationship is established by combining the terrain characteristics with the solar radiation path, and the spatial shading characteristics are calculated by the surface inclination, orientation angle and shading angle, effectively capturing the impact of the terrain on the illumination distribution, reflecting the interference of the landform difference on the photovoltaic output, integrating the meteorological and spatial collaborative characteristics, improving the prediction accuracy while taking into account the nonlinear errors caused by the terrain and spatial distribution, enhancing the adaptability of the power prediction to dynamic disturbances, and realizing the steady-state operation guarantee of the photovoltaic power station cluster under complex terrain conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power prediction, and in particular to a photovoltaic power station cluster power prediction method and system based on multi-task learning. Background Art

[0002] The field of photovoltaic power forecasting involves the advance prediction of photovoltaic power station output power, primarily used in core areas such as power system scheduling, operational optimization, and renewable energy integration. This technology encompasses key issues such as single-station and cluster forecasting for photovoltaic power stations, as well as power uncertainty management under large-scale renewable energy integration. By modeling the output characteristics of photovoltaic power stations affected by weather variations, photovoltaic power forecasting can effectively support the safe and stable operation of power systems, improve the utilization of renewable energy, reduce backup capacity, and enhance the economic efficiency of power grids. This field primarily relies on meteorological data collection and preprocessing, power data analysis, machine learning modeling, and short-term, medium-term, and long-term forecasting algorithms. Core technologies involved include time series-based forecasting methods, physical model-based forecasting methods, and hybrid forecasting methods that integrate meteorological forecasts with power data. A multi-task learning-based photovoltaic power station cluster power forecasting method addresses the spatial correlation and meteorological differences between multiple photovoltaic power stations. By constructing a multi-task learning framework, the power forecasting task for each station is treated as an independent but related subtask. During model training, some feature representations are shared while retaining their specific characteristics, improving overall forecast accuracy. This method utilizes a multi-task modeling approach based on deep neural networks. It extracts common features through shared network layers and captures personalized information about each power plant using independent task-specific network layers. Inputs primarily include site power data, corresponding meteorological observations, and numerical weather forecast data. This allows for simultaneous learning of multiple subtasks, addressing the problem of traditional single-task prediction methods that ignore inter-plant collaboration.

[0003] In scenarios with drastic terrain changes or complex occlusion between arrays, traditional methods fail to incorporate the occlusion factor of the three-dimensional spatial topography on the solar incident path, resulting in the prediction model being unable to accurately identify the lighting differences caused by terrain height differences and orientation differences. Especially in mountainous areas or slope environments with densely distributed power stations, sunshine resource estimation deviations often occur. Existing methods do not involve the temporal evolution analysis of the propagation path of wind disturbances between different power stations, resulting in the problem of asynchronous response time between multiple power stations. The prediction model lacks a reasonable limiting mechanism for output changes under the influence of wind disturbances, resulting in sudden deviations between actual output and predicted values, making it difficult to ensure the prediction's ability to respond to sudden environmental disturbances. In severe cases, it will have an adverse impact on the stability of power grid scheduling. Summary of the Invention

[0004] In order to solve the problem that the existing technology fails to include the factor of occlusion of the sun's incident path by three-dimensional spatial topography, resulting in the prediction model being unable to accurately identify the illumination differences caused by terrain height differences and orientation differences, especially in mountainous areas or slopes with densely distributed power stations, sunshine resource estimation deviations often occur. The existing method does not involve the temporal evolution analysis of the propagation path of wind disturbances between different power stations, resulting in the problem of asynchronous response time between multiple power stations. The prediction model lacks a reasonable limiting mechanism for output changes under the influence of wind disturbances, resulting in sudden deviations between actual output and predicted values, making it difficult to ensure the prediction's ability to respond to sudden environmental disturbances, and in severe cases, it will have an adverse effect on the stability of power grid dispatching. The present invention provides a photovoltaic power station cluster power prediction method and system based on multi-task learning. The technical solution is as follows:

[0005] On the one hand, a photovoltaic power station cluster power prediction method based on multi-task learning is provided, which includes:

[0006] S1: Obtain the three-dimensional spatial vector data of the PV power station's geographic location, detect the slope inclination angle, slope orientation angle, and terrain obstacle shielding angle corresponding to the spatial location point, and construct the solar incidence path vector by combining the solar altitude angle and azimuth angle. Output the terrain shielding spatial feature input item.

[0007] S2: Obtaining the boundary contour vector information of the photovoltaic power station cluster array through the terrain occlusion spatial feature input item, detecting the long side orientation angle, short side orientation angle and tilt angle of the array, calculating the boundary space distance value between the arrays, and generating the array space collaborative feature set;

[0008] S3: Using the array spatial collaborative feature set, mapping the solar incident path vector to the spatial position of the photovoltaic power station, identifying the spatial illumination distribution characteristics under the solar incident path, and generating spatial illumination correlation features;

[0009] S4: Based on the spatial illumination correlation characteristics, the real-time wind speed change information of the photovoltaic power station is identified, the wind speed increment, fluctuation frequency and disturbance duration are extracted, the spatial propagation path of the wind disturbance is constructed, and the time sequence of the disturbance arriving at the photovoltaic power station is derived. By identifying the immediate response, short-time response and delayed response power stations, a disturbance dynamic prediction limit label set is formed.

[0010] As a further solution of the present invention, the terrain occlusion spatial feature input items include slope occlusion angle, ground object occlusion intensity, and solar path intersection density; the array space collaborative feature set includes boundary direction features, spatial occlusion matrix, and array spacing index; the spatial lighting association features include illumination distribution map, spatial lighting gradient, and occlusion attenuation factor; the disturbance dynamic prediction limiting label set includes response delay classification, disturbance propagation trajectory, and limiting adjustment index.

[0011] As a further solution of the present invention, the step of obtaining the terrain occlusion spatial feature input item is specifically as follows:

[0012] S101: Obtain three-dimensional spatial vector data of the photovoltaic power station's geographic location, extract the elevation value of the slope grid corresponding to the spatial point, calculate the elevation change rate of adjacent points in the north-south and east-west directions, and generate the slope azimuth angle value and slope angle value based on the horizontal distance;

[0013] S102: Based on the slope azimuth angle value and the slope angle value, the solar altitude angle and the solar azimuth angle of the spatial position point are called, the slope inclination angle and the slope orientation angle are used as spatial surface feature parameters, and the sun angle is used as the illumination vector parameter. The slope undulation and the occlusion range limit of the terrain occlusion structure are analyzed to generate a terrain obstacle occlusion angle value.

[0014] S103: Based on the terrain obstacle occlusion angle value, perform spatial intersection judgment on the sun's incident path direction vector and the terrain boundary vector, filter obstacle position points with an intersection relationship, and record the corresponding occlusion angle interval to generate a terrain occlusion spatial feature input item.

[0015] As a further solution of the present invention, the steps of acquiring the array space collaborative feature set are specifically as follows:

[0016] S201: Based on the terrain obstruction spatial feature input item, obtaining the boundary outline vector information of the photovoltaic power station cluster array, extracting the boundary vertex coordinate sequence of each array group, calculating the direction vector of the line connecting the starting and ending points in the boundary, and measuring the angle between the line and the true north direction vector to generate the array orientation and boundary angle value;

[0017] S202: Based on the array orientation and boundary angle value, the coordinates of multiple array center points in the array boundary contour vector are called, and the straight-line distance between any two array center points is calculated as the boundary space distance value. At the same time, point pairs whose distance is less than the array boundary error distance range are eliminated to obtain array boundary space distance data;

[0018] The boundary space distance value adopts the formula:

[0019] ;

[0020] in, The array number is and The boundary space distance value of the center point, The array number is The coordinate value of the center point in the X-axis direction, The array number is The coordinate value of the center point in the X-axis direction, The array number is The coordinate value of the center point in the Y-axis direction, The array number is The coordinate value of the center point in the Y-axis direction, The number representing the array boundary angle combination condition is The array boundary angle difference under ;

[0021] S203: Based on the array boundary space distance data, call the solar altitude angle and azimuth angle parameters corresponding to the moment, project the solar vector onto a two-dimensional plane to construct a solar projection path, calculate the angle with the direction vector of the line connecting the center of the array boundary, mark the occlusion relationship of the array pairs whose angles are within the overlapping range of the solar projection path, and generate an array space collaborative feature set.

[0022] As a further solution of the present invention, the steps of acquiring the spatial illumination correlation features are specifically as follows:

[0023] S301: Based on the array spatial collaborative feature set, the solar altitude angle and azimuth angle of the array at the corresponding time point are called to construct a solar incident path vector. According to the position of the array in geographic space, the solar incident path vector is mapped to the center point position of the array within the photovoltaic power station area. The angle between the incident vector and the array orientation vector and the incident direction are recorded to generate a solar path spatial projection angle value.

[0024] S302: Based on the spatial projection angle value of the solar path, the occlusion angle and occlusion position point information in the terrain occlusion spatial feature input item are called, and the overlap of the solar incident path vector and the corresponding terrain occlusion direction vector is judged. The path corresponding to the occlusion point is eliminated to generate a spatial lighting correlation feature.

[0025] As a further solution of the present invention, the steps of obtaining the disturbance dynamic prediction clipping label set are specifically as follows:

[0026] S401: Based on the spatial illumination correlation characteristics, real-time wind speed monitoring data within the photovoltaic power station area is obtained, a wind speed value change curve of each location point is extracted in time series, an incremental characteristic value of wind speed change is calculated, and the frequency of occurrence of continuous rising or falling segments in the speed change is counted to generate a wind speed disturbance sequence feature;

[0027] S402: Based on the wind speed disturbance sequence characteristics, the geographical location of the photovoltaic power station and the terrain shielding spatial characteristics input items are called to construct an initial wind disturbance propagation path, the path is adjusted according to the terrain shielding direction vector, and the propagation distance and propagation time of the disturbance corresponding to the path point are recorded to generate a disturbance propagation time series;

[0028] S403: Based on the disturbance propagation time series, the time node when the power station receives the disturbance signal is compared with the start time of the wind speed disturbance, and the power stations are divided into immediate response, short-time response and delayed response power stations. The response time interval and the corresponding change amplitude are used as label input to generate a disturbance dynamic prediction limit label set.

[0029] As a further solution of the present invention, the wind speed change increment characteristic value adopts the formula:

[0030] ;

[0031] in, Representative location point The incremental characteristic value of wind speed change at Indicates the time series The wind speed value at a moment, and Represents the first Hedi The wind speed value at a moment, Indicates the The spatial illumination weight coefficient of each wind speed point, Indicates the The wind speed value at a moment, represents the average value of wind speed in the selected time series, represents the sum of squared deviations of wind speed values, Indicates the number of wind speed value samples.

[0032] As a further solution of the present invention, the method further includes step S5:

[0033] S5: Based on the disturbance dynamic prediction limit label set, combined with meteorological observation data and real-time irradiation trends of multiple power plants as multi-dimensional inputs of the multi-task learning framework, site power prediction is performed in the task association space, and the prediction output is dynamically adjusted to generate dynamic spatial prediction results for the photovoltaic power plant cluster;

[0034] The dynamic spatial prediction result of the photovoltaic power station cluster includes a site power curve set, a prediction error interval, and a dynamic correction factor.

[0035] As a further solution of the present invention, the steps for obtaining the dynamic spatial prediction results of the photovoltaic power station cluster are specifically as follows:

[0036] S501: Based on the disturbance dynamic prediction limit label set, calling the meteorological observation data corresponding to the photovoltaic power station at the time, collecting the real-time irradiation trend curve of the array area, arranging and combining multiple types of parameters according to time series and spatial position, and generating multi-station meteorological and irradiation joint input characteristic values;

[0037] S502: Based on the joint input feature values ​​of meteorological and irradiation data from multiple power stations and the response category information in the disturbance dynamic prediction clipping label set, differentiated sites are used as associated task nodes in a multi-task learning structure. Power prediction is performed on the task nodes based on the input features and response mode distribution to generate multi-site prediction output values ​​under task association.

[0038] S503: Based on the multi-site prediction output value under the task association, call the disturbance response type and limit range value of the site within the differentiated time window, determine whether the real-time prediction value exceeds the change boundary corresponding to the type, perform truncation correction on the prediction result that exceeds the range, and generate the dynamic spatial prediction result of the photovoltaic power station cluster.

[0039] On the other hand, the photovoltaic power station cluster power prediction system based on multi-task learning is used to execute the photovoltaic power station cluster power prediction method based on multi-task learning, and the system includes:

[0040] The terrain occlusion extraction module obtains three-dimensional spatial vector data of the area where the photovoltaic power station is located, detects the slope inclination angle, slope orientation angle, and slope top elevation value of the geographical location point, and constructs the solar incidence path vector by combining the solar altitude angle and azimuth angle. It then selects grid cells with consistent directions and elevation occlusion to generate terrain occlusion spatial feature input items.

[0041] The array feature recognition module obtains the boundary vector data of the photovoltaic array based on the terrain obstruction spatial feature input item, detects the long side direction, short side direction angle and tilt angle of the array, measures the boundary space distance between arrays, and generates an array space collaborative feature set;

[0042] The illumination association module calls the array spatial collaborative feature set, calls the terrain occlusion angle value and the array tilt angle, determines whether the path angle and the azimuth difference value meet the illumination conditions, analyzes the intersection of the path and the array in the time series, and generates spatial illumination association features;

[0043] The disturbance path derivation module identifies wind speed change data based on the spatial illumination correlation characteristics, extracts wind speed increments, fluctuation frequency, and duration in continuous time, measures the distance between the disturbance propagation path along the main wind direction and the obstruction curve, and generates a disturbance dynamic prediction clipping label set;

[0044] The power prediction output module obtains the real-time meteorological observation data and irradiation trend value of the site through the disturbance dynamic prediction limit label set, maps them to the task subspace in response order, performs power prediction according to the subspace characteristics, and generates the dynamic spatial prediction results of the photovoltaic power station cluster.

[0045] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0046] By combining terrain features with solar radiation paths, a three-dimensional spatial shielding relationship is established. The spatial shielding characteristics are calculated through the surface inclination, orientation angle and shielding angle, effectively capturing the impact of terrain on light distribution and reflecting the interference of landform differences on photovoltaic output. The mutual shielding between arrays is evaluated by using the relationship between array contour vectors and spatial layout. The actual illumination range is derived by combining the sun's altitude and azimuth to achieve spatial cascade shielding modeling. The disturbance response time series under different landform conditions is characterized by the wind speed change trend and the disturbance propagation path, and the response characteristics of each power station under real-time changes are identified. The meteorological and spatial collaborative characteristics are integrated to achieve responsive adjustment to the joint impact of light and disturbance, improve the prediction accuracy while taking into account the nonlinear errors caused by terrain and spatial distribution, improve the sensitivity and timeliness of collaborative prediction among multiple sites, enhance the adaptability of power prediction to dynamic disturbances, and ensure the steady-state operation of photovoltaic power station clusters under complex terrain conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the workflow of the present invention;

[0048] Figure 2 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0049] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0050] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0051] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0052] See also Figure 1 The embodiment of the present invention provides a photovoltaic power station cluster power prediction method based on multi-task learning. The processing flow of the method may include the following steps:

[0053] S1: Obtain the three-dimensional spatial vector data of the PV power station's geographic location, detect the slope inclination angle, slope orientation angle, and terrain obstacle shielding angle corresponding to the spatial location point, construct the solar incidence path vector by combining the solar altitude angle and azimuth angle, determine the intersection relationship between the solar incidence path vector and terrain obstacles, and output the terrain shielding spatial feature input item;

[0054] S2: Using terrain occlusion spatial feature inputs, we obtain the boundary outline vector information of the PV power station cluster arrays, detect the long side orientation angle, short side orientation angle, and tilt angle of the arrays, calculate the boundary spatial distance between arrays, and identify the illumination occlusion relationship between arrays by combining the solar altitude angle and azimuth spatial projection to generate the array spatial collaborative feature set.

[0055] S3: Using the array spatial collaborative feature set, the solar incident path vector is mapped to the spatial location of the photovoltaic power station. Combined with the influence of terrain shading, the spatial illumination distribution characteristics under the solar incident path are identified to generate spatial illumination correlation features.

[0056] S4: Based on spatial illumination correlation characteristics, the real-time wind speed change information of the photovoltaic power station is identified, and the wind speed increment, fluctuation frequency and disturbance duration are extracted. Combined with the geographical location of the power station and terrain shielding input, the spatial propagation path of the wind disturbance is constructed, and the time sequence of the disturbance reaching the photovoltaic power station is derived. By identifying the power stations with immediate response, short-time response and delayed response, the corresponding prediction output change constraints are set to form a disturbance dynamic prediction limit label set;

[0057] S5: Based on the perturbation dynamic prediction clipping label set, combined with meteorological observation data from multiple power plants and real-time irradiation trends as the multi-dimensional input of the multi-task learning framework, site power prediction is performed in the task association space, and the prediction output is dynamically adjusted to generate dynamic spatial prediction results for the PV power plant cluster;

[0058] The input items of terrain shading spatial features include slope shading angle, ground object shading intensity, and solar path intersection density. The array space collaborative feature set includes boundary direction features, spatial shading matrix, and array spacing index. The spatial illumination related features include illumination distribution map, spatial illumination gradient, and shading attenuation factor. The disturbance dynamic prediction limit label set includes response delay classification, disturbance propagation trajectory, and limit adjustment index. The dynamic spatial prediction results of photovoltaic power station clusters include site power curve set, prediction error range, and dynamic correction factor.

[0059] The steps for obtaining terrain occlusion spatial feature input items are as follows:

[0060] S101: Obtain three-dimensional spatial vector data of the photovoltaic power station's geographic location, extract the elevation value of the slope grid corresponding to the spatial point, calculate the elevation change rate of adjacent points in the north-south and east-west directions, and generate the slope azimuth angle value and slope angle value based on the horizontal distance;

[0061] High-precision latitude, longitude, and elevation information of the power station area is obtained through drone aerial surveys or remote sensing satellites. Photovoltaic panels are deployed in a certain mountainous area. First, point coordinate data is obtained through drone flight. The data is converted into a three-dimensional vector data format using geographic information tools and saved as GeoJSON or Shapefile. The point data is matched with the corresponding digital elevation model, and the actual elevation value of each spatial point in the slope grid is extracted. On this basis, an adjacency relationship is established for each point and its adjacent points in the east, west, south, and north directions. The elevation difference between adjacent points and the horizontal distance between the two points are calculated, and the elevation change rate in the north-south direction and the east-west direction is extracted. If the elevation difference between two points with a spacing of 10 meters is 1 meter, the change rate is 0.1. The change rate is combined with the spatial vector analysis method to gradually calculate the slope angle value and the slope azimuth angle value, forming slope angle feature data including slope direction and slope. The slope direction of a point is set to southeast and the slope angle is 5 degrees, and the slope azimuth angle value and slope angle value are generated.

[0062] S102: Based on the slope azimuth angle value and the slope angle value, the solar altitude angle and the solar azimuth angle of the spatial location point are called, the slope inclination angle and the slope orientation angle are used as spatial surface feature parameters, and the sun angle is used as the illumination vector parameter. The slope undulation and the occlusion range limit of the terrain occlusion structure are analyzed to generate the terrain obstacle occlusion angle value;

[0063] Obtain the solar altitude and azimuth angles for the spatial location at the corresponding time. This data can be obtained by calling an existing solar position calculation. Using the NREL solar position library, after inputting the specific time, date, and location coordinates, the program outputs the current solar spatial angle. Given the slope angle and solar angle parameters, the program treats the slope as an inclined plane and the solar incidence direction as a vector. Compare the angle difference between the two and calculate whether the point is within the direct sunlight range or is blocked by terrain structures. This process analyzes the blocking relationship based on the actual terrain conditions. For example, in a hilly area, adjacent high mountains may block the lower slopes. The program scans the terrain in 10° or 5° increments. By traversing each angle direction, it detects whether there is an obstruction in that direction and records the blocking limit angle. A 360-degree blocking angle array is constructed, recording the upper limit of the blocking range in each direction. For example, there is no obstruction in the south direction, and the west direction is blocked by a mountain to the west by 12°. This provides a basis for subsequent determination of whether sunlight is blocked and generates a terrain obstacle blocking angle value.

[0064] S103: Based on the terrain obstacle occlusion angle value, perform spatial intersection judgment on the sun's incident path direction vector and the terrain boundary vector, select obstacle locations with an intersection relationship, and record the corresponding occlusion angle interval to generate terrain occlusion spatial feature input items;

[0065] It is necessary to determine whether the sun's incident path at a specific moment will have an occlusion relationship with the terrain boundary. The specific process is to convert the sun's incident direction into a spatial vector form, then extract the boundary points of the terrain and convert them into a vector set. By comparing the angle between the two vectors, it is determined whether the sun's path geometrically intersects with the terrain boundary. Set it at a certain place and time, the sun's incident direction is close to the southwest direction, and there is a mountain boundary vector in this direction that intersects the sun vector within a certain angle range. Record the intersection point and confirm it as an occlusion point, and simultaneously record the occlusion range of the point. For example, the occlusion range of the point covers an azimuth between 260° and 280°. After traversing the boundary points and completing the screening of all occlusion relationships, the point set with occlusion relationship is structured and stored as an occlusion space feature data item, including the geographical location of each occlusion point, the corresponding occlusion angle range and the maximum occlusion angle value, to generate a terrain occlusion space feature input item.

[0066] The specific steps for obtaining the array space collaborative feature set are:

[0067] S201: Based on the terrain occlusion spatial feature input, obtain the boundary outline vector information of the photovoltaic power station cluster array, extract the boundary vertex coordinate sequence of each array group, calculate the direction vector of the line connecting the starting and ending points in the boundary, and measure the angle between it and the true north direction vector to generate the array orientation and boundary angle value;

[0068] High-resolution remote sensing images or drone aerial images are used to identify the boundary shapes of each photovoltaic array within a power plant. Image processing methods are used to extract the array boundary contours, and vectorization operations are used to generate contour vector data formats, such as polygon-formatted GeoJSON files. In actual deployment scenarios, photovoltaic power plants may have multiple regularly or irregularly arranged array areas. Contour line tracing methods are used to extract the boundaries of each array group and obtain their vertex coordinate sequences. A rectangular array is set to consist of four corner points, recorded as coordinate sequences P1 to P4. A pair of first and last points is extracted from each coordinate sequence, and a start-end line direction vector is constructed to represent the overall orientation of the array. The angle between this vector and a reference vector formed by the standard true north direction is measured. The angle between the two direction vectors is compared using the angle measurement module in GIS. The start and end vectors of a certain array are set to form a 45° angle with the true north direction. The orientation angle of the array is recorded as 45°, and the array orientation and boundary angle values ​​are generated.

[0069] S202: Based on the array orientation and boundary angle value, the coordinates of multiple array center points in the array boundary contour vector are called, and the straight-line distance between any two array center points is calculated as the boundary space distance value. At the same time, point pairs whose distance is less than the array boundary error distance range are eliminated to obtain array boundary space distance data;

[0070] The boundary space distance value is calculated using the formula:

[0071] ;

[0072] in, The array number is and The boundary space distance value of the center point, The array number is The coordinate value of the center point in the X-axis direction, The array number is The coordinate value of the center point in the X-axis direction, The array number is The coordinate value of the center point in the Y-axis direction, The array number is The coordinate value of the center point in the Y-axis direction, The number representing the array boundary angle combination condition is The array boundary angle difference under ;

[0073] Formula calculation logic: The comprehensive spatial distance is formed by calculating the sum of the squared distances between the center points of the two arrays in the X and Y directions, and adding a weighted correction term introduced by the array boundary angle difference. This correction term is based on the product of the mean of the horizontal coordinate of the center point and the difference between the boundary angles. It is combined with the Y-axis difference to form an absolute difference, which is used to reflect the offset effect between the two points when the array direction changes. The angle difference is calculated from the direction-finding sensor sampling data and corrected by normalization and fluctuation adjustment coefficient to enhance the sensitivity of the angle response. The resulting value is compared with the set error threshold to filter out point pairs that do not meet the spatial separation requirements.

[0074] The boundary space distance refers to the spatial separation between the center points of two arrays, including the influence of the orientation angle compensation. This distance integrates the geometric differences of the arrays under the changes of plane coordinates and boundary directions.

[0075] Parameter meaning and formula calculation derivation process:

[0076] The spatial distribution of the array center point is acquired in real time using a laser ranging sensor, with coordinate accuracy controlled within a ±2 mm error range. Coordinate data is derived from the calibration database of the automated array positioning system. The array's orientation angle toward the boundary is acquired using a high-precision gyroscope array unit direction-finding sensor sampling frequency of 100 Hz. The sampled data undergoes a bidirectional weighted sliding average to determine the angle variation range and output the boundary angle difference.

[0077] The coordinates of the center point of array i are rice, meters, the coordinates of the center point of array j are rice, rice;

[0078] The difference in the array boundary angle combination is Radians. This angle value is extracted by processing the difference between the maximum and minimum angles obtained by the direction-finding sensor within 90 seconds using a standard normalization model. The direction-finding accuracy is corrected using the matching sensor calibration coefficient. The sensor calibration deviation is corrected to 0.002 radians. The normalization standard range is 0 to π. The angle difference is calculated as follows:

[0079] ;

[0080] in, radian, Radians, degrees normalization coefficient:

[0081] ;

[0082] in, 、 All of them are derived from the initial input data of the direction-finding sensor, and we get:

[0083] ;

[0084] In order to control the overall angle difference within the boundary sensitivity threshold, an adjustment coefficient is introduced (This coefficient changes with the maximum fluctuation range of direction finding. When the fluctuation range exceeds 0.1 radian, Increase by 0.01, confirmed by the 10-day angle fluctuation statistics), and corrected to:

[0085] ;

[0086] Substitute into the main formula and calculate:

[0087] Calculate the squared difference part:

[0088] ;

[0089] ;

[0090] Calculate the angle correction term:

[0091] ;

[0092] ;

[0093] ;

[0094] Combine the above three items to calculate:

[0095] ;

[0096] The result shows that the boundary space distance value between array i and array j under the array boundary angle combination condition c is 4.824 meters. This value is logically associated with the subsequent elimination of point pairs that are smaller than the boundary error threshold. If the boundary error distance threshold is set to 3.5 meters, the point pair is retained and enters the next step of the spatial boundary configuration calculation logic process.

[0097] S203: Based on the array boundary spatial distance data, the solar altitude angle and azimuth parameters corresponding to the time are called, the solar vector is projected onto a two-dimensional plane to construct a solar projection path, and the angle between the solar vector and the direction vector of the line connecting the array boundary centers is calculated. Array pairs whose angles are within the overlapping range of the solar projection path are marked as having an occlusion relationship, and an array spatial collaborative feature set is generated;

[0098] Solar radiation parameters are introduced for spatial occlusion judgment. The solar altitude and azimuth parameters at a specified time are called upon. A standard solar position algorithm is used to calculate the solar incident direction at that moment. This three-dimensional incident direction vector is projected onto a two-dimensional plane, retaining only the horizontal component of the ground surface. This solar projection path is constructed. Assuming the sun is southeast at 10 a.m., the projection path is deflected 135°. The direction vector of the line connecting the array center points obtained above is extracted and compared with the angle between them. For point pairs with an angle within a preset range, set to less than 20°, an occlusion probability is determined and an occlusion relationship is marked. Array point pairs with such geometric relationships are summarized. This feature set not only contains the spatial geometric relationship between arrays, but also carries attribute fields such as the center point coordinates, line direction, angle value, and whether the marked point pair overlaps with the solar projection path. In actual PV power plant operation, this feature set can be used to analyze potential occlusion impacts between arrays and adjust the layout design accordingly, generating an array spatial collaboration feature set.

[0099] The specific steps for obtaining spatial illumination correlation features are as follows:

[0100] S301: Based on the array spatial collaborative feature set, the solar altitude angle and azimuth angle at the corresponding time point of the array are called to construct a solar incident path vector. According to the array's position in geographic space, the solar incident path vector is mapped to the array center point position within the photovoltaic power station area. The angle between the incident vector and the array orientation vector and the incident direction are recorded to generate the solar path spatial projection angle value.

[0101] Call the solar altitude angle and solar azimuth angle at the time point corresponding to the array, and use the solar position calculation algorithm to obtain the specific position of the sun in the sky at the current moment. Set it at 9:00 on June 21, 2025. For a photovoltaic power station located at 35° north latitude, the solar altitude angle is 45° and the azimuth angle is 110°. Construct the solar incident path vector based on the angle value and regard it as a vector representation of the three-dimensional direction. Map the incident vector according to the geographic space position of the array. The center point of each array is used as the projection target point, and execute each center point in turn. The angle between the incident path and the array orientation is calculated. The array orientation is set to 95°, the current sun incidence direction is 110°, and the angle between the two is recorded as 15°. At the same time, it is marked that the current sun enters the array surface from the southeast direction, and the incidence direction is southeast. An array is formed with each array as a unit, which contains fields such as array number, time point, incidence angle, sun direction, angle with array orientation, and other information. This array can be used as the basic input data for subsequent analysis of the relationship between illumination and occlusion, and is used to match terrain occlusion conditions and further eliminate affected areas to generate the spatial projection angle value of the sun path.

[0102] S302: Based on the spatial projection angle value of the sun path, the occlusion angle and occlusion position point information in the terrain occlusion spatial feature input item are called, and the overlap between the sun incident path vector and the corresponding terrain occlusion direction vector is determined. The path corresponding to the occlusion point is eliminated to generate a spatial illumination correlation feature;

[0103] Call the occlusion angles and corresponding occlusion point information in each direction recorded in the terrain occlusion spatial feature input item, and perform the overlap judgment of the sun's incident path and the terrain occlusion direction. During the specific operation, extract the two-dimensional projection direction corresponding to the sun's incident vector at a certain time point, set the current sun projection direction to 125°, and then retrieve the corresponding occlusion point direction information from the terrain occlusion feature set to determine whether the direction contains the incident angle. Set the occlusion data to record the existence of occlusion points in the direction of 120° to 130°, and the maximum occlusion angle is 15°. At this time, it is confirmed that the sun's path overlaps with the occlusion direction, and at the same time, it is proposed The spatial position coordinates and angle range corresponding to the occlusion point are taken out, the path is marked as an occlusion path, and it is removed from the valid projection path. The above judgment and elimination process is performed on the projection paths in multiple directions in sequence, and the unobstructed paths and occlusion paths are identified and classified to form a spatial illumination-related feature dataset. It records whether there is a visible illumination channel at each array center point at a specific time. The accompanying attribute fields include array number, occlusion status, occlusion angle range, actual illumination incident direction and other information, which provide data support for the subsequent judgment of whether the array effectively receives solar radiation and generate spatial illumination-related features.

[0104] The specific steps for obtaining the perturbation dynamic prediction clipping label set are as follows:

[0105] S401: Based on spatial illumination correlation characteristics, real-time wind speed monitoring data within the photovoltaic power station area is obtained, and wind speed value change curves of location points are extracted in time series. The incremental characteristic value of wind speed change is calculated, and the frequency of occurrence of continuous rising or falling segments in the speed change is counted to generate wind speed disturbance sequence characteristics.

[0106] The incremental characteristic value of wind speed change is calculated using the formula:

[0107] ;

[0108] in, Representative location point The incremental characteristic value of wind speed change at Indicates the time series The wind speed value at a moment, and Represents the first Hedi The wind speed value at a moment, Indicates the The spatial illumination weight coefficient of each wind speed point, Indicates the The wind speed value at a moment, represents the average value of wind speed in the selected time series, represents the sum of squared deviations of wind speed values, Indicates the number of wind speed value samples;

[0109] The formula is used to calculate the incremental characteristic value of wind speed change at a certain location in a time series. The formula consists of three parts: the first term represents the wind speed difference between the current location and the next moment, reflecting the short-term wind speed trend; the second term is the product of the wind speed deviation of the sample point and the light weight, quantifying the impact of spatial disturbance on wind speed change, where the light intensity is normalized to the weight value; the third term is the absolute value of the wind speed difference between the current location and the previous moment multiplied by the square root of the sum of the squares of the overall deviation, reflecting the combined effect of the change and the overall fluctuation characteristics. The absolute value of the three terms is added and divided by the number of samples to obtain the disturbance intensity after unit balance, which reflects the wind speed change characteristics at the corresponding location.

[0110] The incremental eigenvalue of wind speed change indicates the strength of wind speed change at a certain moment. It is a value calculated by combining time series trend, spatial disturbance and fluctuation mutation. This eigenvalue reflects the sensitivity of wind speed change and can be used to identify anomalies and trends in disturbance sequences.

[0111] Parameter meaning and formula calculation derivation process:

[0112] Select a monitoring point time series wind speed value sample number n=10, collect real-time wind speed value They are: 6.5, 6.8, 7.0, 6.9, 7.2, 7.1, 7.3, 7.0, 6.9, 7.0 (unit: meters per second), and the arithmetic mean wind speed is calculated as:

[0113] ;

[0114] Spatial illumination weight coefficient of monitoring point The weight value is obtained by quantization of the light correlation model and normalized according to the real-time light intensity. … They are: 0.12, 0.15, 0.10, 0.13, 0.14, 0.11, 0.12, 0.13, 0.14, 0.16, and the weight values ​​are adjusted with the fluctuation of light intensity; the sum of squared sample deviations is given by ;

[0115] Calculated

[0116] ;

[0117] Focus on the position point a=5, then Substituting into the formula:

[0118] ;

[0119]

[0120] ;

[0121] Bring in:

[0122] ;

[0123] The results show that the incremental characteristic value of wind speed change is 0.03664, which can be used for subsequent frequency statistics and disturbance sequence generation. It is obtained by the combined effect of the wind speed difference at the location point, the light weighted deviation term, and the sum of squared deviation terms.

[0124] S402: Based on the wind speed disturbance sequence characteristics, the geographical location of the photovoltaic power station and the terrain shielding spatial characteristics are called upon to construct an initial wind disturbance propagation path. The path is adjusted according to the terrain shielding direction vector, and the propagation distance and propagation time of the disturbance corresponding to the path point are recorded to generate a disturbance propagation time series.

[0125] Combining the geographic location of the photovoltaic power station with the spatial feature input of terrain shielding, the initial propagation path of the disturbance is constructed with the starting location of the disturbance as the initial propagation source. The path is dominated by the wind speed direction. The initial vector direction is determined according to the dominant wind direction at the time of the disturbance. The wind direction is set to southwest at the beginning of the disturbance, and the propagation path is constructed with the southwest direction as the starting direction. The direction vectors of each position in the terrain shielding spatial data are then introduced to adjust the propagation path. The adjustment method is to check whether the path coincides with or is close to the shielding direction. If a terrain mountain with a shielding direction between 230° and 250° is encountered in a certain path direction, the propagation path is redirected to bypass the shielding direction. The disturbance transmission efficiency is corrected according to the degree of path deviation. The propagation distance of each position point in the adjusted path and the time point of the disturbance signal arrival are recorded. The time required for propagation from the disturbance source to an array 300 meters away is set to 32 seconds. The propagation trajectory and time consumption of the array position point under different disturbances are calculated in sequence to generate a disturbance propagation time series.

[0126] S403: Based on the disturbance propagation time series, the time nodes when the power station receives the disturbance signal are compared with the start time of the wind speed disturbance, and the power stations are divided into immediate response, short-time response and delayed response power stations. The response time interval and the corresponding change amplitude are used as label input to generate the disturbance dynamic prediction limit label set;

[0127] By comparing the actual time nodes when each array receives the disturbance signal with the disturbance start time, the response time difference of each array is calculated. According to the time difference, the photovoltaic array is divided into three categories: immediate response, short-time response and delayed response. Immediate response refers to the array that responds within 3 seconds after the signal arrives, short-time response refers to the array that responds between 3 and 10 seconds, and delayed response refers to the array that responds after more than 10 seconds. After the classification is completed, the change amplitude of the corresponding disturbance segment is extracted from the wind speed disturbance sequence, and the multidimensional label is constructed in combination with the response time. Array A is set to be an immediate response, and its wind speed change is +1.2m / s. Array B is set to be a delayed response, and its wind speed change is , the response time and disturbance amplitude are taken as label input to form a disturbance dynamic prediction limit label set. By constructing an array structure, the label information is uniformly managed according to the array number. Each label data contains fields such as response classification, response time interval, disturbance amplitude, propagation distance, wind direction angle, etc., to generate a disturbance dynamic prediction limit label set.

[0128] The specific steps for obtaining the dynamic spatial prediction results of a photovoltaic power station cluster are as follows:

[0129] S501: Based on the disturbance dynamic prediction limit label set, the meteorological observation data corresponding to the photovoltaic power station is called, the real-time irradiation trend curve of the array area is collected, and multiple types of parameters are arranged and combined according to time series and spatial position to generate the multi-station meteorological and irradiation joint input characteristic value;

[0130] The meteorological observation data corresponding to the current moment of the photovoltaic power station is called, which usually includes temperature, humidity, wind speed, atmospheric pressure and cloud cover changes, and on this basis, the real-time solar irradiation trend curve of each array area is collected. The irradiation data comes from the light intensity sensors installed between the arrays or is estimated from satellite cloud images. It is set that within a certain period of time, the irradiation intensity of array A drops from 800W / m² to 600W / m². The downward trend is recorded and time-matched with factors such as temperature increase and wind speed mutation. The meteorological data and irradiation trend data are aligned in chronological order, with minutes as the time unit. Multiple types of parameters are arranged point by point according to their positions in space, and the time axis is unified to construct an input data structure based on the "time-space" dual axis. At the same time, the response category field in the disturbance label set is introduced and matched with the meteorological-irradiation combination corresponding to each array. Array A is set to belong to the "short-term response" and its meteorological input characteristics are high wind speed, high humidity, and reduced irradiation. This combination is uniformly classified into the multi-station meteorological and irradiation joint input eigenvalue structure as the input source of the subsequent power prediction model. After completing data cleaning, alignment and label mapping, the multi-station meteorological and irradiation joint input eigenvalue is generated.

[0131] S502: Based on the joint input feature values ​​of meteorological and irradiation data from multiple power stations and the response category information in the disturbance dynamic prediction limit label set, differentiated sites are used as associated task nodes in the multi-task learning structure. Power prediction is performed on the task nodes based on the input features and response pattern distribution to generate multi-site prediction output values ​​under task association.

[0132] The response category information in the disturbance dynamic prediction limit label set is introduced into the multi-task learning structure, and photovoltaic power station arrays in different response ranges are classified and processed. Sites with obvious response differences are used as independent task nodes in the multi-task learning model. The "immediate response" site and the "delayed response" site are set as two subtasks. The input feature distribution and response relationship of each task node are processed separately. By establishing a mapping relationship between input features and power output at each node, a "short-term response" task node is set. The model receives inputs such as wind speed, radiation intensity decrease rate, and temperature fluctuation, and predicts that the power change trend of the site at the next moment will decrease by 50kW. After the prediction is completed at the task node, the output of each node is summarized to form the multi-site prediction output value under task association.

[0133] S503: Based on the multi-site prediction output values ​​under task association, the disturbance response type and limit range value of the site within the differentiated time window are called to determine whether the real-time prediction value exceeds the change boundary corresponding to the type, and the prediction result that exceeds the range is truncated to generate the dynamic spatial prediction result of the photovoltaic power station cluster;

[0134] Combined with the response type and limit range defined in the disturbance label set for each array, the response characteristics recorded by the site in the differentiated time window are called, and a certain array is set to "delayed response". Its allowed variation after the disturbance is triggered is ±30kW. An over-limit judgment is performed on the predicted output value at the current moment. If the power prediction value of a certain array is found to be higher than the set upper limit, it is judged to exceed the type response boundary. The prediction of array B is set to a decrease of 75kW. Since this array belongs to the "short-term response" category and the maximum allowed variation is 50kW, its output is automatically truncated to The corrected prediction results are marked and saved, and the truncated correction operation is completed for all sites in turn to ensure that the overall predicted output value fluctuates within the acceptable range of disturbance. It not only considers the input characteristics and disturbance response of each array, but also reflects the disturbance variation range that the site can withstand in actual application. It includes fields such as site number, predicted value, correction status, response type, change boundary, etc., providing a stable input reference for scheduling and control, and forming a dynamic spatial prediction result for the photovoltaic power station cluster.

[0135] See also Figure 2 , a photovoltaic power station cluster power prediction system based on multi-task learning, the system includes:

[0136] The terrain occlusion extraction module obtains three-dimensional spatial vector data of the area where the photovoltaic power station is located, detects the slope inclination angle, slope orientation angle, and slope top elevation value of the geographical location point, and constructs the solar incidence path vector by combining the solar altitude angle and azimuth angle. It then selects grid cells with consistent directions and elevation occlusion to generate terrain occlusion spatial feature input items.

[0137] The array feature recognition module obtains the boundary vector data of the photovoltaic array based on the terrain occlusion spatial feature input, detects the long side direction, short side direction angle and tilt angle of the array, measures the boundary space distance between arrays, and generates the array space collaborative feature set;

[0138] The illumination correlation module calls the array spatial collaborative feature set, the terrain occlusion angle value and the array tilt angle, determines whether the path angle and azimuth difference value meet the illumination conditions, analyzes the intersection of the path and the array in the time series, and generates spatial illumination correlation features;

[0139] The disturbance path derivation module identifies wind speed change data based on spatial illumination correlation characteristics, extracts wind speed increments, fluctuation frequency, and duration over continuous time, measures the distance between the disturbance propagation path along the main wind direction and the obstruction curve, and generates a disturbance dynamic prediction limit label set.

[0140] The power prediction output module obtains the real-time meteorological observation data and irradiation trend value of the site by perturbing the dynamic prediction limit label set, maps them to the task subspace in response order, performs power prediction according to the subspace characteristics, and generates the dynamic spatial prediction results of the photovoltaic power station cluster.

[0141] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A photovoltaic power station cluster power prediction method based on multi-task learning, characterized by: The following steps are involved: S1: Obtain the three-dimensional spatial vector data of the PV power station's geographic location, detect the slope inclination angle, slope orientation angle, and terrain obstacle shielding angle corresponding to the spatial location point, and construct the solar incidence path vector by combining the solar altitude angle and azimuth angle. Output the terrain shielding spatial feature input item. S2: Obtaining the boundary contour vector information of the photovoltaic power station cluster array through the terrain occlusion spatial feature input item, detecting the long side orientation angle, short side orientation angle and tilt angle of the array, calculating the boundary space distance value between the arrays, and generating the array space collaborative feature set; S3: Using the array spatial collaborative feature set, mapping the solar incident path vector to the spatial position of the photovoltaic power station, identifying the spatial illumination distribution characteristics under the solar incident path, and generating spatial illumination correlation features; S4: Based on the spatial illumination correlation characteristics, the real-time wind speed change information of the photovoltaic power station is identified, the wind speed increment, fluctuation frequency and disturbance duration are extracted, the spatial propagation path of the wind disturbance is constructed, and the time sequence of the disturbance reaching the photovoltaic power station is derived. By identifying the immediate response, short-time response and delayed response power stations, a disturbance dynamic prediction limit label set is formed; S5: Based on the disturbance dynamic prediction limit label set, combined with meteorological observation data and real-time irradiation trends of multiple power plants as multi-dimensional inputs of the multi-task learning framework, site power prediction is performed in the task association space, and the prediction output is dynamically adjusted to generate dynamic spatial prediction results for the photovoltaic power plant cluster; The dynamic spatial prediction results of the photovoltaic power station cluster include a site power curve set, a prediction error interval, and a dynamic correction factor; The terrain occlusion spatial feature input items include slope occlusion angle, ground object occlusion intensity, and solar path intersection density; the array space collaborative feature set includes boundary direction features, spatial occlusion matrix, and array spacing index; the spatial illumination association features include illumination distribution map, spatial illumination gradient, and occlusion attenuation factor; the disturbance dynamic prediction limiting label set includes response delay classification, disturbance propagation trajectory, and limiting adjustment index.

2. The photovoltaic power station cluster power prediction method based on multi-task learning according to claim 1 is characterized in that: The steps for obtaining the terrain occlusion spatial feature input item are specifically as follows: S101: Obtain three-dimensional spatial vector data of the photovoltaic power station's geographic location, extract the elevation value of the slope grid corresponding to the spatial point, calculate the elevation change rate of adjacent points in the north-south and east-west directions, and generate the slope azimuth angle value and slope angle value based on the horizontal distance; S102: Based on the slope azimuth angle value and the slope angle value, the solar altitude angle and the solar azimuth angle of the spatial position point are called, the slope inclination angle and the slope orientation angle are used as spatial surface feature parameters, and the sun angle is used as the illumination vector parameter. The slope undulation and the occlusion range limit of the terrain occlusion structure are analyzed to generate a terrain obstacle occlusion angle value. S103: Based on the terrain obstacle occlusion angle value, perform spatial intersection judgment on the sun's incident path direction vector and the terrain boundary vector, filter obstacle position points with an intersection relationship, and record the corresponding occlusion angle interval to generate a terrain occlusion spatial feature input item.

3. The photovoltaic power station cluster power prediction method based on multi-task learning according to claim 2 is characterized in that: The steps for obtaining the array space collaborative feature set are specifically as follows: S201: Based on the terrain obstruction spatial feature input item, obtaining the boundary outline vector information of the photovoltaic power station cluster array, extracting the boundary vertex coordinate sequence of each array group, calculating the direction vector of the line connecting the starting and ending points in the boundary, and measuring the angle between the line and the true north direction vector to generate the array orientation and boundary angle value; S202: Based on the array orientation and boundary angle value, the coordinates of multiple array center points in the array boundary contour vector are called, and the straight-line distance between any two array center points is calculated as the boundary space distance value. At the same time, point pairs whose distance is less than the array boundary error distance range are eliminated to obtain array boundary space distance data; The boundary space distance value adopts the formula: ； in, The array number is and The boundary space distance value of the center point, The array number is The coordinate value of the center point in the X-axis direction, The array number is The coordinate value of the center point in the X-axis direction, The array number is The coordinate value of the center point in the Y-axis direction, The array number is The coordinate value of the center point in the Y-axis direction, The number representing the array boundary angle combination condition is The array boundary angle difference under ; S203: Based on the array boundary space distance data, call the solar altitude angle and azimuth angle parameters corresponding to the moment, project the solar vector onto a two-dimensional plane to construct a solar projection path, calculate the angle with the direction vector of the line connecting the center of the array boundary, mark the occlusion relationship of the array pairs whose angles are within the overlapping range of the solar projection path, and generate an array space collaborative feature set.

4. The photovoltaic power station cluster power prediction method based on multi-task learning according to claim 3 is characterized in that: The steps for obtaining the spatial illumination correlation features are specifically as follows: S301: Based on the array spatial collaborative feature set, the solar altitude angle and azimuth angle of the array at the corresponding time point are called to construct a solar incident path vector. According to the position of the array in geographic space, the solar incident path vector is mapped to the center point position of the array within the photovoltaic power station area. The angle between the incident vector and the array orientation vector and the incident direction are recorded to generate a solar path spatial projection angle value. S302: Based on the spatial projection angle value of the solar path, the occlusion angle and occlusion position point information in the terrain occlusion spatial feature input item are called, and the overlap of the solar incident path vector and the corresponding terrain occlusion direction vector is judged. The path corresponding to the occlusion point is eliminated to generate a spatial lighting correlation feature.

5. The photovoltaic power station cluster power prediction method based on multi-task learning according to claim 4 is characterized in that: The steps for obtaining the disturbance dynamic prediction clipping label set are specifically as follows: S401: Based on the spatial illumination correlation characteristics, real-time wind speed monitoring data within the photovoltaic power station area is obtained, a wind speed value change curve of each location point is extracted in time series, an incremental characteristic value of wind speed change is calculated, and the frequency of occurrence of continuous rising or falling segments in the speed change is counted to generate a wind speed disturbance sequence feature; S402: Based on the wind speed disturbance sequence characteristics, the geographical location of the photovoltaic power station and the terrain shielding spatial characteristics input items are called to construct an initial wind disturbance propagation path, the path is adjusted according to the terrain shielding direction vector, and the propagation distance and propagation time of the disturbance corresponding to the path point are recorded to generate a disturbance propagation time series; S403: Based on the disturbance propagation time series, the time node when the power station receives the disturbance signal is compared with the start time of the wind speed disturbance, and the power stations are divided into immediate response, short-time response and delayed response power stations. The response time interval and the corresponding change amplitude are used as label input to generate a disturbance dynamic prediction limit label set.

6. The photovoltaic power station cluster power prediction method based on multi-task learning according to claim 5 is characterized in that: The wind speed change increment characteristic value adopts the formula: ； in, Representative location point The incremental characteristic value of wind speed change at Indicates the time series The wind speed value at a moment, and Represents the first Hedi The wind speed value at a moment, Indicates the The spatial illumination weight coefficient of each wind speed point, Indicates the The wind speed value at a moment, represents the average value of wind speed in the selected time series, represents the sum of squared deviations of wind speed values, Indicates the number of wind speed value samples.

7. The photovoltaic power station cluster power prediction method based on multi-task learning according to claim 1 is characterized in that: The steps for obtaining the dynamic spatial prediction results of the photovoltaic power station cluster are specifically as follows: S501: Based on the disturbance dynamic prediction limit label set, calling the meteorological observation data corresponding to the photovoltaic power station at the time, collecting the real-time irradiation trend curve of the array area, arranging and combining multiple types of parameters according to time series and spatial position, and generating multi-station meteorological and irradiation joint input characteristic values; S502: Based on the joint input characteristic values ​​of meteorological and irradiation data from multiple power stations and the response category information in the disturbance dynamic prediction limit label set, differentiated sites are used as associated task nodes in a multi-task learning structure, power prediction is performed on the task nodes based on the input characteristics and response patterns, and multi-site prediction output values ​​under task association are generated; S503: Based on the multi-site prediction output value under the task association, call the disturbance response type and limit range value of the site within the differentiated time window, determine whether the real-time prediction value exceeds the change boundary corresponding to the type, perform truncation correction on the prediction result that exceeds the range, and generate the dynamic spatial prediction result of the photovoltaic power station cluster.

8. Photovoltaic power station cluster power prediction system based on multi-task learning, characterized by: The system is used to implement the photovoltaic power station cluster power prediction method based on multi-task learning according to any one of claims 1 to 7, and the system includes: The terrain occlusion extraction module obtains three-dimensional spatial vector data of the area where the photovoltaic power station is located, detects the slope inclination angle, slope orientation angle, and slope top elevation value of the geographical location point, and constructs the solar incidence path vector by combining the solar altitude angle and azimuth angle. It then selects grid cells with consistent directions and elevation occlusion to generate terrain occlusion spatial feature input items. The array feature recognition module obtains the boundary vector data of the photovoltaic array based on the terrain obstruction spatial feature input item, detects the long side orientation angle, short side orientation angle and tilt angle of the array, measures the boundary space distance between arrays, and generates an array space collaborative feature set; The illumination association module calls the array spatial collaborative feature set, calls the terrain occlusion angle value and the array tilt angle, determines whether the path angle and the azimuth difference value meet the illumination conditions, analyzes the intersection of the path and the array in the time series, and generates spatial illumination association features; The disturbance path derivation module identifies wind speed change data based on the spatial illumination correlation characteristics, extracts wind speed increments, fluctuation frequency, and disturbance duration in continuous time, measures the distance between the disturbance propagation path along the main wind direction and the occlusion curve, and generates a disturbance dynamic prediction limit label set; The power prediction output module obtains the real-time meteorological observation data and irradiation trend value of the site through the disturbance dynamic prediction limit label set, maps them to the task subspace in response order, performs power prediction according to the subspace characteristics, and generates the dynamic spatial prediction results of the photovoltaic power station cluster.

Images