A solar power generation power prediction method based on big data analysis
By using multi-source photovoltaic preprocessing and feature transformation, combined with an improved SOFTS network, the problem of insufficient characterization of the state relationship of multi-source data in solar power prediction is solved, and more accurate and stable power prediction is achieved.
Patent Information
- Application Number
- CN202611310791.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-10
AI Technical Summary
Existing solar power generation prediction methods are insufficient in characterizing the state relationships between multi-source photovoltaic data, making it difficult to distinguish the coupling trajectory between irradiance changes and power response. Furthermore, they neglect the sequential relationship between cloud cover, irradiance, module temperature, and inverter status, resulting in insufficient prediction stability.
A photovoltaic state fusion feature is constructed by employing multi-source photovoltaic preprocessing, irradiance-power phase diagram trajectory convolution transformation of the MultiROCKET algorithm, and cross-variable homogeneous response permutation coding. Combined with an improved SOFTS network, solar power generation is predicted, generating state switching core representation, station core representation, and thermal inertia state embedding.
It improves the stability and accuracy of solar power generation prediction, can accurately identify irradiance decline, recovery and abnormal fluctuation states, and enhances the characterization of equipment-level feature fusion and component temperature hysteresis response.
Smart Images

Figure CN123073722A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data analytics, and in particular to a method for predicting solar power generation based on big data analytics. Background Technology
[0002] With the continuous expansion of new energy power generation, the integration ratio of solar power plants into the power system is constantly increasing. Solar power generation forecasting has become a crucial technical aspect in grid dispatching, power planning, energy storage coordination control, and plant operation management. Solar power generation is influenced by factors such as irradiance, cloud cover, ambient temperature, module temperature, wind speed and direction, humidity, inverter operating status, module layout, and plant installed capacity, exhibiting diurnal variations, weather disturbances, and equipment status changes. Existing solar power generation forecasting methods typically collect historical power generation data and meteorological data to establish photovoltaic power forecasting models, then output forecasts for short-term or ultra-short-term power generation. With the development of big data analytics and artificial intelligence forecasting technologies, Long Short-Term Memory (LSTM) networks, convolutional neural networks, time series forecasting models, and ensemble learning methods have been applied to solar power generation forecasting scenarios, extracting power generation patterns from historical power curves and meteorological variables.
[0003] Current methods for predicting solar power generation still suffer from insufficient characterization of the state relationships between multi-source photovoltaic data. Most methods concatenate historical power, irradiance, cloud cover, module temperature, and inverter status as ordinary multivariate sequences, lacking a dedicated extraction of the coupling trajectory between irradiance changes and power response. This makes it difficult to distinguish between states such as declining irradiance with a lagging power decline, irradiance recovery with insufficient power recovery, and stable irradiance with abnormal power fluctuations. Existing methods also typically ignore the sequential relationship between cloud cover, irradiance, module temperature, inverter status, and power sequences, leading to confusion between weather-driven power fluctuations, insufficient power recovery due to module thermal inertia, and power anomalies caused by equipment operating conditions. Furthermore, traditional prediction models often use a uniform time window for feature fusion, making it difficult to reorganize the sequence before and after anchor points near photovoltaic state switching candidate points. They also struggle to generate site-level prediction representations based on the equipment hierarchy relationships between module regions, inverters, and the site. The lag effect of module temperature changes on power output also lacks an independent memory structure, resulting in insufficient prediction stability under scenarios such as cloud shadow entry, cloud cover recovery, power ramp-up, and changes in the state of local inverter clusters.
[0004] Therefore, how to provide a method for predicting solar power generation based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a solar power prediction method based on big data analysis. This invention utilizes multi-source photovoltaic (PV) prediction data preprocessing, the MultiROCKET algorithm's irradiance-power phase diagram trajectory convolutional transformation and cross-variable homogeneous response permutation encoding, PV state fusion feature construction, and improved SOFTS network prediction technology. It details the implementation process of multi-source data acquisition from solar power plants, PV state feature extraction, state switching core representation generation, plant core representation generation, thermal inertia state embedding generation, and power prediction result output. This achieves joint modeling of irradiance response relationships, cross-variable response sequences, PV state switching, inverter cluster hierarchy relationships, and component thermal inertia effects in solar power changes. Compared to traditional prediction methods that directly stitch together multi-source time series, this invention offers advantages such as more complete expression of PV state changes, more accurate cloud shadow disturbance identification, more complete fusion of equipment hierarchy features, clearer characterization of component temperature hysteresis response, and higher stability in solar power prediction.
[0006] A solar power generation prediction method based on big data analysis according to an embodiment of the present invention includes: Multi-source photovoltaic (PV) prediction data from solar power plants is collected and preprocessed to generate a standardized PV prediction dataset. Based on this dataset, a multi-source PV time series is constructed, and a connection table for plant equipment is generated according to the connection relationships between component regions, inverters, and the plant. The multi-source PV time series is input into the MultiROCKET algorithm to perform irradiance-power phase diagram trajectory convolution transformation and cross-variable co-core response permutation coding, generating an irradiance-power response trajectory feature matrix, a cross-variable co-core response permutation coding table, and PV state switching candidate points. Anchor point segmentation is performed on the multi-source PV time series based on these candidate points, and a PV state fusion feature set is generated by combining the plant equipment connection table, the irradiance-power response trajectory feature matrix, and the cross-variable co-core response permutation coding table. An improved SOWTS network, comprising a PV state anchor point core refresh layer, an inverter cluster-level core folding layer, and a component thermal inertia memory slot layer, is constructed. This network is then input into the PV state fusion feature set to generate a state switching core representation, a plant core representation, and a thermal inertia state embedding. Based on the state switching core representation, the station core representation, and the thermal inertia state embedding, the solar power generation prediction results are generated; actual power generation feedback data are collected, prediction feedback records are generated, and the standardized photovoltaic prediction dataset, MultiROCKET algorithm parameters, and improved SOFTS network parameters are updated according to the prediction feedback records.
[0007] Optionally, generating the standardized photovoltaic prediction dataset includes: Collect multi-source photovoltaic forecast data from solar power plants. The multi-source photovoltaic forecast data includes historical power generation data, irradiance data, cloud cover data, ambient temperature data, module temperature data, wind speed and direction data, humidity data, inverter status data, module regional operation data, plant installed capacity data, and weather forecast data. The multi-source photovoltaic prediction data are correlated according to the collection time, site number, inverter number, component area number and data type identifier to generate a photovoltaic raw prediction data sequence. The original photovoltaic prediction data sequence is processed by deleting duplicate records, removing invalid records, handling abnormal jump values, and filling missing values to generate a cleaned photovoltaic prediction data sequence. The cleaned photovoltaic prediction data sequence is processed by timestamp alignment, sampling interval unification, unit unification and numerical standardization to generate a standardized photovoltaic prediction data sequence. The standardized photovoltaic forecast data sequence is archived according to the site number, inverter number, module area number and collection time to generate a standardized photovoltaic forecast dataset.
[0008] Optionally, the construction of the multi-source photovoltaic time series and the generation of the power station equipment connection table include: Read the site number, inverter number, module area number, acquisition time and data type identifier from the standardized photovoltaic prediction dataset, and generate a photovoltaic data index table; According to the photovoltaic data index table, historical power generation data, irradiance data, cloud cover data, ambient temperature data, module temperature data, wind speed and direction data, humidity data, and inverter status data are written into the corresponding time positions to generate a multi-source photovoltaic time series. The multi-source photovoltaic time series were arranged according to the collection time sequence, and then divided into power time series, irradiance time series, cloud cover time series, ambient temperature time series, module temperature time series, wind speed and direction time series, humidity time series, and inverter status time series according to the data type identifier. Read the connection relationship data between the module area, inverter and the station in the solar power plant, and establish the correspondence between the module area number and the inverter number, and the correspondence between the inverter number and the station number; The component area number, inverter number, site number, installed capacity data, and connection level identifier are associated to generate a site equipment connection relationship table.
[0009] Optionally, the generation of the irradiance-power response trajectory feature matrix, the transvariable homonuclear response permutation coding table, and photovoltaic state switching candidate points includes: Read the power time series, irradiance time series, cloud cover time series, module temperature time series, wind speed and direction time series, and inverter status time series from the multi-source photovoltaic time series, and construct the MultiROCKET input window according to the acquisition time order; The power change and irradiance change at adjacent acquisition times are calculated for the power time series and irradiance time series respectively. The irradiance change and power change at the same acquisition time are combined into irradiance-power phase diagram points, and the irradiance-power phase diagram trajectory sequence is generated according to the acquisition time sequence. The irradiation-power phase diagram trajectory sequence is subjected to irradiation-power phase diagram trajectory convolution transformation, that is, the irradiation-power phase diagram trajectory sequence is converted into a two-channel trajectory sequence. The convolution response is calculated on the two-channel trajectory sequence through the convolution kernel set in the MultiROCKET algorithm, and the irradiation-power response trajectory feature matrix is generated by positive proportional pooling, positive mean pooling, positive index mean pooling and longest continuous positive pooling. Perform cross-variable homogeneous kernel response permutation coding on power time series, irradiance time series, cloud cover time series, component temperature time series, wind speed and direction time series and inverter state time series. Apply the same set of convolution kernels to each time series and record the response position and response order of the same convolution kernel in different time series. The cross-variable response results of the same convolution kernel are arranged according to the convolution kernel number, time series type, response position and response order to generate a cross-variable same kernel response arrangement coding table; Based on the convolution response sign switching position in the irradiance-power response trajectory feature matrix, the boundary of the longest continuous positive value interval, and the response arrangement change position in the cross-variable homogeneous response arrangement coding table, photovoltaic state switching candidate points are generated according to the acquisition time.
[0010] Optionally, the generation of the photovoltaic state fusion feature set includes: Read the multi-source photovoltaic time series, photovoltaic state switching candidate points, station equipment connection relationship table, irradiance-power response trajectory feature matrix and cross-variable homogeneous response permutation coding table; The photovoltaic state switching candidate points are sorted according to the acquisition time sequence, and the photovoltaic state switching candidate points are mapped to the acquisition time positions in the multi-source photovoltaic time series to generate a photovoltaic state anchor point index table. Anchor point segmentation is performed on the multi-source photovoltaic time series according to the photovoltaic state anchor point index table to generate an anchor point segmented sequence including the pre-anchor point sequence, the post-anchor point sequence, and the cross-anchor point sequence; Based on the site equipment connection relationship table, the anchor point segment sequence is associated with the corresponding component area number, inverter number and site number respectively, and an anchor point segment equipment association table is generated. The irradiation-power response trajectory feature matrix and the transvariable homonuclear response arrangement coding table are written into the anchor point segmentation equipment association table according to the acquisition time and position to generate the anchor point status feature table. Based on the anchor point state characteristic table, photovoltaic state anchor point index table, site equipment connection relationship table, and component temperature time series in multi-source photovoltaic time series, a photovoltaic state fusion feature set is generated, which includes anchor point segment sequence, photovoltaic state switching candidate points, inverter cluster hierarchical relationship data, component thermal state data, irradiance-power response trajectory characteristics, and cross-variable co-core response permutation coding.
[0011] Optionally, the generation of the state switching core representation, the station core representation, and the thermal inertial state embedding includes: An improved SOFTS network is constructed, which includes a photovoltaic state anchor core refresh layer, an inverter cluster-level core folding layer, and a module thermal inertia memory slot layer. The photovoltaic state anchor core refresh layer reads the anchor segment sequence and photovoltaic state switching candidate points, generates the core representation before the anchor point, the core representation after the anchor point, and the core representation across anchor points, respectively, and refreshes the core representation across anchor points according to the acquisition time position corresponding to the photovoltaic state switching candidate point to generate the state switching core representation; The inverter cluster hierarchical core folding layer reads the inverter cluster hierarchical relationship data and generates the component area core representation, inverter cluster core representation and site core representation according to the hierarchical relationship from the component area to the inverter and from the inverter to the site. The component thermal inertia memory slot layer reads the component thermal state data, establishes component heating memory slots, component cooling memory slots, irradiation recovery hysteresis memory slots, and temperature rise power decay memory slots, and generates thermal inertia state embeddings according to the time sequence of each memory slot. The improved SOFTS network was trained using a combination of power prediction error, state switching core representation consistency error, inverter cluster hierarchical folding error, and thermal inertia state embedding error as joint optimization objectives. The network parameters of the photovoltaic state anchor core refresh layer, inverter cluster hierarchical core folding layer, and module thermal inertia memory slot layer were continuously optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved SOFTS network was considered to have completed convergence training.
[0012] Optionally, generating the solar power generation prediction result includes: Read the state switching core representation, the station core representation, the thermal inertia state embedding, and the photovoltaic state switching candidate points, and generate a prediction time step identifier according to the prediction task, and generate a prediction input representation sequence according to the prediction time step identifier; Map the state switching core representation to the state switching prediction vector, map the station core representation to the station power prediction vector, and map the thermal inertial state embedding to the thermal inertial correction vector. According to the prediction time step order, the state switching prediction vector, the station power prediction vector, and the thermal inertia correction vector are matched to generate the prediction time step power representation; The power prediction sequence for the power plant is generated based on the predicted time step power representation, and the power prediction sequence for the inverter cluster is generated based on the power plant equipment connection table. The power prediction sequence of the power plant, the power prediction sequence of the inverter cluster, the prediction time step identifier, and the photovoltaic state switching candidate points are combined to generate the solar power generation prediction result.
[0013] Optionally, the step of collecting and updating the actual power generation feedback data includes: Collect actual power generation feedback data, actual meteorological feedback data, component temperature feedback data, and inverter status feedback data within the forecast period; Based on the site number, inverter number, component area number, acquisition time, and prediction time step identifier, the actual power generation feedback data and the solar power generation prediction results are matched to generate a prediction result correspondence table. Based on the prediction result correspondence table, the difference between the actual power generation and the corresponding predicted power value is calculated, a prediction error sequence is generated, and the prediction error sequence, actual meteorological feedback data, component temperature feedback data and inverter status feedback data are combined to generate a prediction feedback record. The standardized photovoltaic prediction dataset is updated based on the prediction feedback records, and the convolution kernel set parameters, pooling statistics parameters, and response permutation encoding parameters of the MultiROCKET algorithm are updated based on the prediction error sequence. Based on the prediction feedback records, update and improve the parameters of the photovoltaic state anchor core refresh layer, inverter cluster level core fold layer, and module thermal inertia memory slot layer of the SOFTS network.
[0014] The beneficial effects of this invention are: This invention proposes a solar power generation prediction method based on big data analysis. By constructing multi-source photovoltaic prediction data processing, multi-source photovoltaic time series construction, MultiROCKET algorithm feature transformation, photovoltaic state fusion feature construction, and an improved SOFTS network prediction process, it enhances the ability to express the state relationships of multi-source data in solar power generation prediction. Compared to the traditional method of directly concatenating historical power, irradiance, cloud cover, module temperature, and inverter state into the prediction model, this invention utilizes irradiance-power phase diagram trajectory convolutional transformation to extract the coupling relationship between irradiance changes and power response. This allows for the differentiation of irradiance decline, irradiance recovery, power lag response, and abnormal power fluctuation states, improving feature recognition capabilities under cloud shadow disturbance and power ramp-up scenarios.
[0015] This invention utilizes cross-variable homogeneous response permutation coding to perform response sequence analysis on power time series, irradiance time series, cloud cover time series, module temperature time series, wind speed and direction time series, and inverter state time series. This expresses the sequential relationship between meteorological changes, module temperature changes, equipment state changes, and power changes, reducing confusion between meteorological-driven power fluctuations, insufficient power recovery due to module thermal inertia, and power changes caused by equipment state anomalies. Simultaneously, this invention performs anchor point segmentation on multi-source photovoltaic time series based on photovoltaic state switching candidate points and generates a photovoltaic state fusion feature set by combining it with a site equipment connection relationship table, improving prediction stability during state switching phases and in scenarios with local equipment fluctuations.
[0016] This invention constructs an improved SOWTS network comprising a photovoltaic state anchor core refresh layer, an inverter cluster-level core folding layer, and a module thermal inertia memory slot layer. This generates state switching core representations, site core representations, and thermal inertia state embeddings, respectively, enabling the solar power generation prediction process to simultaneously consider photovoltaic state switching, site equipment hierarchy relationships, and the impact of module temperature lag. Compared to traditional unified time window prediction models, this invention can generate prediction representations that better reflect actual operating conditions under scenarios such as cloud shadow entry, cloud cover recovery, rapid power changes, local state changes in inverter clusters, and module temperature rise and degradation. Furthermore, it continuously updates relevant data and parameters based on actual power generation feedback data, improving the accuracy, continuity, and operational adaptability of solar power generation prediction results. Attached Figure Description
[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a solar power generation prediction method based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram illustrating the process of performing irradiance-power phase diagram trajectory convolution transformation and cross-variable homo-kernel response permutation encoding using the MultiROCKET algorithm of the solar power prediction method based on big data analysis proposed in this invention. Figure 3 This is a schematic diagram of the structure of an improved SOFTS network for a solar power generation prediction method based on big data analysis proposed in this invention. Detailed Implementation
[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0019] refer to Figure 1, Figure 2 and Figure 3 A method for predicting solar power generation based on big data analysis, comprising: Multi-source photovoltaic (PV) prediction data from solar power plants is collected and preprocessed to generate a standardized PV prediction dataset. Based on this dataset, a multi-source PV time series is constructed, and a connection table for plant equipment is generated according to the connection relationships between component regions, inverters, and the plant. The multi-source PV time series is input into the MultiROCKET algorithm to perform irradiance-power phase diagram trajectory convolution transformation and cross-variable co-core response permutation coding, generating an irradiance-power response trajectory feature matrix, a cross-variable co-core response permutation coding table, and PV state switching candidate points. Anchor point segmentation is performed on the multi-source PV time series based on these candidate points, and a PV state fusion feature set is generated by combining the plant equipment connection table, the irradiance-power response trajectory feature matrix, and the cross-variable co-core response permutation coding table. An improved SOWTS network, comprising a PV state anchor point core refresh layer, an inverter cluster-level core folding layer, and a component thermal inertia memory slot layer, is constructed. This network is then input into the PV state fusion feature set to generate a state switching core representation, a plant core representation, and a thermal inertia state embedding. Based on the state switching core representation, the station core representation, and the thermal inertia state embedding, the solar power generation prediction results are generated; actual power generation feedback data are collected, prediction feedback records are generated, and the standardized photovoltaic prediction dataset, MultiROCKET algorithm parameters, and improved SOFTS network parameters are updated according to the prediction feedback records.
[0020] In this embodiment, generating the standardized photovoltaic prediction dataset includes: Collect multi-source photovoltaic forecast data from solar power plants. The multi-source photovoltaic forecast data includes historical power generation data, irradiance data, cloud cover data, ambient temperature data, module temperature data, wind speed and direction data, humidity data, inverter status data, module regional operation data, plant installed capacity data, and weather forecast data. The multi-source photovoltaic prediction data are correlated according to the collection time, site number, inverter number, component area number and data type identifier to generate a photovoltaic raw prediction data sequence. The original photovoltaic prediction data sequence is processed by deleting duplicate records, removing invalid records, handling abnormal jump values, and imputing missing values to generate a cleaned photovoltaic prediction data sequence, wherein: The invalid record removal and abnormal transition value handling are performed as follows: Read the acquisition time, acquisition object number, data type identifier and data value from the original photovoltaic prediction data sequence, and establish the valid value range of power data, irradiance data, module temperature data, ambient temperature data, wind speed and direction data and inverter status data according to the data type identifier; Data records that exceed the valid value range are marked as invalid, where: Records with power data less than 0, power data exceeding 1.05 times the corresponding installed capacity of the site, or inverters in shutdown status with corresponding power data not equal to 0 are marked as invalid power records. Records with irradiance data less than 0, irradiance data exceeding 1400W / m², or solar altitude angle not greater than 0 and irradiance data exceeding 50W / m² are marked as invalid irradiance records. Records with component temperature data less than -40℃, component temperature data exceeding 90℃, ambient temperature data less than -50℃, ambient temperature data exceeding 60℃, wind speed data less than 0m / s, wind speed data exceeding 60m / s, wind direction data less than 0°, or wind direction data not less than 360° are marked as invalid meteorological records. Records with an inverter status identifier that is empty or whose inverter status identifier does not belong to the set of operating, standby, power limited, fault, and shutdown status are marked as invalid status records. Invalid power records, invalid irradiance records, invalid meteorological records, and invalid status records are deleted, and the remaining data records are rearranged according to the acquisition time order. Perform abnormal jump value processing on the remaining data records and calculate the amount of data change of the same data object between adjacent collection times; When the power change exceeds 20% of the rated power of the corresponding inverter, and the irradiance change does not exceed 80W / m² and the cloud cover change does not exceed 15% at the same sampling time, the corresponding power record will be marked as an abnormal power jump record. When the irradiance change exceeds 300W / m², and the cloud cover change does not exceed 10% at the same collection time, and the irradiance change direction of adjacent component areas is inconsistent, the corresponding irradiance record will be marked as an abnormal jump irradiance record. When the temperature change of the component exceeds 8℃, and the temperature change of the ambient temperature does not exceed 2℃ and the irradiance change does not exceed 100W / m², the corresponding component temperature record will be marked as an abnormal temperature jump record. The abnormal power jump record, abnormal irradiation jump record, and abnormal temperature jump record are corrected. During the correction, the data values of the previous acquisition time, the next acquisition time, and the corresponding data values of adjacent component areas within the same inverter cluster are read for the same acquisition object. The median of the data values of the previous acquisition time, the next acquisition time, and the corresponding data values of adjacent component areas within the same inverter cluster is written into the acquisition time position corresponding to the abnormal jump record. The cleaned photovoltaic prediction data sequence is processed by timestamp alignment, sampling interval unification, unit unification and numerical standardization to generate a standardized photovoltaic prediction data sequence. The standardized photovoltaic forecast data sequence is archived according to the site number, inverter number, module area number and collection time to generate a standardized photovoltaic forecast dataset.
[0021] In this embodiment, the construction of the multi-source photovoltaic time series and the generation of the power station equipment connection table include: Read the site number, inverter number, module area number, acquisition time and data type identifier from the standardized photovoltaic prediction dataset, and generate a photovoltaic data index table; According to the photovoltaic data index table, historical power generation data, irradiance data, cloud cover data, ambient temperature data, module temperature data, wind speed and direction data, humidity data, and inverter status data are written into the corresponding time positions to generate a multi-source photovoltaic time series. The multi-source photovoltaic time series were arranged according to the collection time sequence, and then divided into power time series, irradiance time series, cloud cover time series, ambient temperature time series, module temperature time series, wind speed and direction time series, humidity time series, and inverter status time series according to the data type identifier. Read the connection relationship data between the module area, inverter, and station in the solar power plant, and establish the correspondence between the module area number and the inverter number, and the correspondence between the inverter number and the station number. Specifically: Read the equipment topology data of the solar power plant. The equipment topology data includes the component area number, inverter number, plant number, and installed capacity. Based on the connection relationship between the component area number and the inverter number, multiple component areas connected to the same inverter number are divided into the same inverter cluster, and an inverter cluster number is generated. Based on the aggregation relationship between inverter numbers and site numbers, establish the correspondence between inverter clusters and site numbers; Associating component area number, inverter cluster number, inverter number, site number, and installed capacity generates a hierarchical relationship between component area, inverter, and site; The component area number, inverter number, site number, installed capacity data, and connection level identifier are associated to generate a site equipment connection relationship table.
[0022] In this embodiment, the generation of the irradiance-power response trajectory feature matrix, the transvariable homogeneous response permutation coding table, and the photovoltaic state switching candidate points includes: Read the power time series, irradiance time series, cloud cover time series, module temperature time series, wind speed and direction time series, and inverter status time series from the multi-source photovoltaic time series, and construct a MultiROCKET input window according to the acquisition time order, where: To construct a MultiROCKET input window, the following steps are taken: Read the acquisition time identifier corresponding to each time series and establish a unified time axis according to the acquisition time order; write the power value, irradiance value, cloud cover value, component temperature value, wind speed and direction value and inverter status value corresponding to the same acquisition time to the same time position to generate multivariate time point records; Time windows are extracted based on 24 consecutive acquisition time positions, and the data values corresponding to each data type within the time window are arranged as input channels. The power input channel, irradiance input channel, cloud cover input channel, module temperature input channel, wind speed and direction input channel, and inverter status input channel are combined to generate the MultiROCKET input window; The time window is moved along a unified time axis at 6 acquisition time positions to generate multiple consecutive MultiROCKET input windows, and the start and end acquisition times of each MultiROCKET input window are recorded. For the power time series and irradiance time series, the power change and irradiance change at adjacent acquisition times are calculated respectively. The irradiance change and power change at the same acquisition time are combined into irradiance-power phase diagram points, and an irradiance-power phase diagram trajectory sequence is generated according to the acquisition time sequence, where: The irradiation-power phase diagram trajectory sequence is generated as follows: Read the power time series and irradiance time series from the MultiROCKET input window, and determine adjacent acquisition times according to the acquisition time sequence; Perform difference calculation on the power values corresponding to every two adjacent acquisition times to generate the power change at the current acquisition time; Perform difference calculation on the irradiance values corresponding to every two adjacent acquisition times to generate the irradiance change at the current acquisition time; The irradiance change at the same acquisition time is taken as the horizontal component, and the power change at the same acquisition time is taken as the vertical component. The points of the irradiance-power phase diagram are generated by combining them. Multiple irradiation-power phase diagram points are arranged sequentially according to the acquisition time to generate an irradiation-power phase diagram trajectory sequence. An irradiation-power phase diagram trajectory convolution transformation is performed on the irradiation-power phase diagram trajectory sequence, converting it into a two-channel trajectory sequence. The convolutional response is then calculated on the two-channel trajectory sequence using the convolutional kernel set in the MultiROCKET algorithm. Finally, an irradiation-power response trajectory feature matrix is generated through positive proportional pooling, positive mean pooling, positive indexed mean pooling, and longest continuous positive pooling. The irradiation-power phase diagram trajectory sequence is converted into a two-channel trajectory sequence, specifically as follows: Read each irradiation-power phase diagram point in the irradiation-power phase diagram trajectory sequence, and extract the irradiation change and power change in each irradiation-power phase diagram point according to the acquisition time sequence; Arrange the irradiance changes corresponding to each collection time in chronological order to generate the first trajectory channel; Arrange the power changes corresponding to each acquisition time in the order of acquisition time to generate the second trajectory channel; Align the first and second trajectory channels according to the same acquisition time position to generate a two-channel trajectory sequence. Each time position in the two-channel trajectory sequence includes the irradiance change and power change corresponding to the acquisition time. The convolution response is calculated for the two-channel trajectory sequence, specifically as follows: The first and second trajectory channels are arranged into a trajectory input matrix according to the acquisition time sequence. The rows of the trajectory input matrix correspond to the acquisition time position, the first column of the trajectory input matrix corresponds to the irradiance change, and the second column of the trajectory input matrix corresponds to the power change. Read the set of convolutional kernels in the MultiROCKET algorithm. Each convolutional kernel in the set includes the kernel number, kernel length, dilation interval, bias term, and kernel element sequence. Convolution kernels are selected sequentially according to their numbers. Based on the kernel length and dilation interval, trajectory segments are extracted from the trajectory input matrix along the acquisition time direction. The trajectory segments include the irradiance change and power change selected according to the dilation interval. The first channel response value is obtained by multiplying the sequence of convolution kernel elements and the irradiance change in the trajectory segment item by item according to the acquisition time position and summing them. The second channel response value is obtained by multiplying the sequence of convolution kernel elements and the power change in the trajectory segment item by item according to the acquisition time position and summing them. Add the response value of the first channel to the response value of the second channel, and subtract the bias term of the corresponding convolution kernel to generate the irradiance-power joint convolution response value of the current convolution kernel at the current acquisition time position; Move the trajectory segment according to the acquisition time direction, repeatedly calculate the irradiance-power joint convolution response value corresponding to each acquisition time position, and generate the irradiance-power trajectory convolution response sequence corresponding to the current convolution kernel. The irradiation-power response trajectory feature matrix is generated as follows: Read the irradiation-power trajectory convolution response sequence corresponding to each convolution kernel, and establish a response sequence index according to the convolution kernel number; For each irradiation-power trajectory convolutional response sequence, count the number of response values greater than 0, and use the ratio of the number of response values greater than 0 to the length of the response sequence as the result of positive proportional pooling. For each irradiation-power trajectory convolution response sequence, sum the response values greater than 0, and use the ratio of the summation result to the number of response values greater than 0 as the result of positive value mean pooling. Record the acquisition time location index of each irradiation-power trajectory convolution response sequence where the response value is greater than 0. Sum the acquisition time location indices and use the ratio of the sum to the number of response values greater than 0 as the result of positive value index mean pooling. The number of consecutive positive response values in each irradiation-power trajectory convolution response sequence is counted according to the acquisition time sequence, and the maximum number of consecutive positive response values is taken as the longest consecutive positive value pooling result. Arrange the positive proportional pooling results, positive mean pooling results, positive index mean pooling results, and longest consecutive positive pooling results corresponding to the same convolution kernel according to the convolution kernel number to generate an irradiation-power response trajectory feature vector. Write the irradiation-power response trajectory feature vectors corresponding to all convolution kernels into the matrix according to the convolution kernel number order to generate the irradiation-power response trajectory feature matrix. Transvariate homogeneous kernel response permutation coding was performed on the power time series, irradiance time series, cloud cover time series, module temperature time series, wind speed and direction time series, and inverter state time series. The same set of convolutional kernels was applied to each time series, and the response position and order of the same convolutional kernel in different time series were recorded. The same set of convolutional kernels is applied to each time series separately, specifically: Select the same convolution kernel sequentially according to the kernel number, and apply the selected kernel to the power time series, irradiance time series, cloud cover time series, component temperature time series, wind speed and direction time series, and inverter status time series respectively. For each time series, the sequence segments are extracted according to the acquisition time order and the convolution kernel dilation interval. The sequence of convolution kernel elements is multiplied with the data values in the sequence segments one by one and summed to generate the convolution response value at the corresponding acquisition time position. Move the sequence segment along the acquisition time direction and repeatedly calculate the convolution response value of the same convolution kernel in the corresponding time series to generate a univariate convolution response sequence; Based on the time series type, associate the univariate convolution response sequences corresponding to the same convolution kernel; The cross-variable response results of the same convolutional kernel are arranged according to the convolutional kernel number, time series type, response position, and response order to generate a cross-variable same-kernel response arrangement coding table, where: The transvariable response results for the same convolution kernel are arranged as follows: Read the univariate convolution response sequences corresponding to the same convolution kernel, and determine the power response sequence, irradiance response sequence, cloud cover response sequence, component temperature response sequence, wind speed and direction response sequence, and inverter status response sequence according to the time series type. In each univariate convolution response sequence, find the acquisition time position where the response value is greater than 0, and record the acquisition time position where the response value is greater than 0 as the response position of the corresponding time sequence; Based on the order of the acquisition time corresponding to the response position, the response positions of the same convolutional kernel in different time series are sorted to generate the cross-variable response order corresponding to the same convolutional kernel; Write the convolution kernel number, time series type, response position, response value, and cross-variable response order into the same encoding line to generate a single convolution kernel response arrangement record; Summarize all single convolution kernel response records in order of convolution kernel number to generate a cross-variable same kernel response arrangement coding table; Based on the convolution response sign switching position in the irradiance-power response trajectory feature matrix, the boundary of the longest continuous positive value interval, and the response arrangement change position in the cross-variable homogeneous response arrangement coding table, photovoltaic state switching candidate points are generated according to the acquisition time, where: Generate photovoltaic state switching candidate points, specifically as follows: Read the irradiation-power trajectory convolution response sequence and find the acquisition time positions where the convolution response value changes from positive to non-positive and from non-positive to positive according to the acquisition time order, and generate the convolution response sign switching position; Read the irradiation-power trajectory convolution response sequence, count the number of consecutive positive response values in the irradiation-power trajectory convolution response sequence according to the acquisition time order, and record the start acquisition time position and end acquisition time position corresponding to each segment of consecutive positive response values. The start acquisition time position and end acquisition time position corresponding to the interval with the largest number of consecutive positive response values are taken as the boundary of the longest consecutive positive value interval. Read the cross-variable co-core response arrangement encoding table, compare the cross-variable response order corresponding to adjacent acquisition time positions, and when the time series type arrangement in the cross-variable response order changes, record the acquisition time position where the arrangement changes as the response arrangement change position. The convolution response sign switching position, the boundary of the longest continuous positive value interval, and the response arrangement change position are merged according to the acquisition time, and duplicate acquisition time positions are deleted to generate a candidate time position set. Based on the candidate time location set, the corresponding acquisition time location is extracted from the multi-source photovoltaic time series, and the corresponding acquisition time location is written into the photovoltaic state switching candidate point.
[0023] In this embodiment, generating the photovoltaic state fusion feature set includes: Read the multi-source photovoltaic time series, photovoltaic state switching candidate points, station equipment connection relationship table, irradiance-power response trajectory feature matrix and cross-variable homogeneous response permutation coding table; The photovoltaic state switching candidate points are sorted according to the acquisition time sequence, and the photovoltaic state switching candidate points are mapped to the acquisition time positions in the multi-source photovoltaic time series to generate a photovoltaic state anchor point index table. Anchor point segmentation is performed on the multi-source photovoltaic time series based on the photovoltaic state anchor point index table, generating an anchor point segmented sequence including the pre-anchor point sequence, the post-anchor point sequence, and the cross-anchor point sequence, where: Generate the anchor point segment sequence, specifically: Read the photovoltaic status switching candidate points, candidate point acquisition time locations, and corresponding multi-source photovoltaic time series numbers from the photovoltaic status anchor point index table; Using the acquisition time position corresponding to each photovoltaic state switching candidate point as the anchor point position, the multi-source photovoltaic time series data corresponding to 6 acquisition time positions are extracted forward along the acquisition time sequence to generate the pre-anchor point sequence. The multi-source photovoltaic time series data corresponding to 6 acquisition time positions are extracted backward along the acquisition time sequence to generate the post-anchor point sequence. The multi-source photovoltaic time series data corresponding to the anchor point position are extracted forward and backward along the acquisition time sequence, while retaining the multi-source photovoltaic time series data corresponding to the anchor point position to generate the cross-anchor point sequence. The pre-anchor sequence, post-anchor sequence, and cross-anchor sequence corresponding to the same photovoltaic state switching candidate point are associated according to the candidate point collection time and location to generate a single anchor point segment record; All single anchor point segment records are summarized in chronological order of collection time to generate an anchor point segment sequence. Based on the site equipment connection relationship table, the anchor point segment sequence is associated with the corresponding component area number, inverter number and site number respectively, and an anchor point segment equipment association table is generated. The irradiation-power response trajectory feature matrix and the transvariable homonuclear response arrangement coding table are written into the anchor point segmentation equipment association table according to the acquisition time and position to generate the anchor point status feature table. Based on the anchor point state characteristic table, the photovoltaic state anchor point index table, the site equipment connection relationship table, and the component temperature time series in the multi-source photovoltaic time series, a photovoltaic state fusion feature set is generated, including anchor point segmentation sequence, photovoltaic state switching candidate points, inverter cluster hierarchical relationship data, component thermal state data, irradiance-power response trajectory characteristics, and cross-variable homogeneous response permutation coding, where: The inverter cluster hierarchy data is generated from the field equipment connection relationship table. The inverter cluster hierarchy data includes component area number, inverter number, inverter cluster number, field number, installed capacity data and connection level identifier. Among them, multiple component areas connected to the same inverter number belong to the same inverter cluster. The component thermal status data is generated from the component temperature time series. The component thermal status data includes the component area number, acquisition time, component temperature value, ambient temperature value, irradiance value and corresponding power value.
[0024] In this embodiment, the generation of the state switching core representation, the station core representation, and the thermal inertia state embedding includes: An improved SOFTS network is constructed, comprising a photovoltaic state anchor core refresh layer, an inverter cluster-level core folding layer, and a module thermal inertia memory slot layer, wherein: The following steps are taken to build an improved SOFTS network: Based on the traditional SOFTS network's input embedding layer, STAR aggregation-redistribution module, and prediction output layer, the structure is modified. The traditional SOFTS network receives multivariate time series through the input embedding layer, generates a core representation of the sequence through the STAR aggregation-redistribution module, and generates time series prediction results through the prediction output layer. In the STAR aggregation-redistribution module, the structure of generating a single core representation based on a complete input window is adjusted to a photovoltaic state anchor core refresh layer. The photovoltaic state anchor core refresh layer receives anchor segment sequences and photovoltaic state switching candidate points and generates a state switching core representation. An inverter cluster-level core folding layer is added between the input embedding layer and the STAR aggregation-redistribution module. The inverter cluster-level core folding layer receives inverter cluster-level relationship data and generates a site core representation. A component thermal inertia memory slot layer is added between the STAR aggregation-redistribution module and the prediction output layer. The component thermal inertia memory slot layer receives component thermal state data and generates a thermal inertia state embedding. The state switching core representation, site core representation, and thermal inertia state embedding are input into the prediction output layer to obtain the improved SOFTS network. The photovoltaic state anchor point core refresh layer reads the anchor point segment sequence and photovoltaic state switching candidate points, and generates the core representation before the anchor point, the core representation after the anchor point, and the cross-anchor point core representation, respectively. It then refreshes the cross-anchor point core representation based on the acquisition time position corresponding to the photovoltaic state switching candidate point, generating the state switching core representation, where: The core refresh layer of the photovoltaic status anchor point includes: Anchor point index register: stores the acquisition time position and candidate point number corresponding to the photovoltaic state switching candidate point; Anchor point segmented cache queue: stores the sequence before anchor point, the sequence after anchor point, and the sequence across anchor points; Core Generation Calculator: Generates core representations before anchor points, core representations after anchor points, and core representations across anchor points; Anchor point location lookup table: records the position number of the photovoltaic state switching candidate point in the cross-anchor point sequence; Core refresh gating: refreshes the cross-anchor core representation based on the anchor point position number; State Switching Core Output: Outputs the state switching core representation; In the core refresh layer of photovoltaic state anchor points, the acquisition time location and candidate point number stored in the anchor point index register are input into the anchor point location lookup table to determine the position number of the photovoltaic state switching candidate point in the cross-anchor point sequence. The anchor point segmented buffer queue inputs the pre-anchor point sequence, post-anchor point sequence, and cross-anchor point sequence into the core generator calculator to generate the pre-anchor point core representation, post-anchor point core representation, and cross-anchor point core representation. The core refresh gate refreshes the cross-anchor point core representation according to the anchor point position number and passes the refreshed cross-anchor point core representation to the state switching core output device to generate the state switching core representation. Generate core representations before anchor points, core representations after anchor points, and core representations across anchor points, specifically as follows: Read the pre-anchor sequence, post-anchor sequence, and cross-anchor sequence from the anchor segment sequence, and establish the correspondence between the three types of sequences according to the photovoltaic state switching candidate point number; Embedding mapping is performed on the time position data in the pre-anchor sequence, post-anchor sequence, and cross-anchor sequence respectively to generate the pre-anchor embedding sequence, post-anchor embedding sequence, and cross-anchor embedding sequence; Channel normalization and time dimension summarization are performed on the pre-anchor embedding sequence, post-anchor embedding sequence, and cross-anchor embedding sequence, respectively. The summarization results are then subjected to dimension transformation to generate the pre-anchor core representation, post-anchor core representation, and cross-anchor core representation. Associate the core representation before anchor point, the core representation after anchor point, and the core representation across anchor points corresponding to the same photovoltaic state switching candidate point number to generate an anchor point core representation group; Refresh the cross-anchor core representation, specifically: Read the acquisition time position corresponding to the anchor point core representation group and the photovoltaic state switching candidate point, and determine the anchor point position number of the photovoltaic state switching candidate point in the cross-anchor point sequence; The cross-anchor point core representation is divided into the core area before the anchor point, the core area at the anchor point location, and the core area after the anchor point according to the anchor point position number. Write the core representation before the anchor point into the core area before the anchor point, write the core representation after the anchor point into the core area after the anchor point, and retain the state change information in the core area of the anchor point position. The core area before the anchor point, the core area at the anchor point location, and the core area after the anchor point are recombined according to the collection time sequence to generate a refreshed cross-anchor point core representation. The refreshed cross-anchor point core representation is archived according to the photovoltaic state switching candidate point number to generate a state switching core representation; The inverter cluster-level core folding layer reads the inverter cluster hierarchical relationship data and generates the module region core representation, inverter cluster core representation, and site core representation according to the hierarchical relationship from module region to inverter and from inverter to site. The inverter cluster-level core folded layer includes: Component Region Index Table: Stores the component region number, the inverter number to which it belongs, and the site number to which it belongs; Inverter cluster relationship register: stores the inverter cluster number and the set of component area numbers corresponding to the inverter cluster number; Component Region Core Cache Queue: Stores the component region embedding sequence and component region core representation according to the component region number; Intra-cluster core folding calculator: Folds the component region core representation under the same inverter cluster into an inverter cluster core representation; Site aggregation table: stores the aggregation relationship between inverter cluster number and site number; Site Core Folding Calculator: Folds the inverter cluster core representation under the same site into a site core representation; Hierarchical core outputs: output component area core representation, inverter cluster core representation, and site core representation; In the inverter cluster-level core folding layer, the component region number, the inverter number to which it belongs, and the site number to which it belongs, recorded in the component region index table, are input into the inverter cluster relationship register to determine the set of component region numbers corresponding to the same inverter cluster. The component region core cache queue stores the component region embedding sequence and the component region core representation according to the component region number, and inputs the component region core representation into the cluster core folding calculator to generate the inverter cluster core representation. The site aggregation relationship table inputs the inverter cluster core representation under the same site into the site core folding calculator according to the aggregation relationship between the inverter cluster number and the site number to generate the site core representation. The hierarchical core output device organizes and outputs the component region core representation, the inverter cluster core representation, and the site core representation. The core representation of the generated component region is as follows: Read the component region number from the inverter cluster hierarchy data, and extract the anchor point segment sequence, component thermal state data, irradiance-power response trajectory features, and cross-variable homocore response permutation code corresponding to the component region from the photovoltaic state fusion feature set based on the component region number; According to the order of acquisition time, the anchor point segment sequence corresponding to the component area is written into the core cache queue of the component area, and the component thermal state data, irradiation-power response trajectory characteristics and cross-variable same-core response arrangement code at the corresponding acquisition time position are written into the same acquisition time position. Perform embedding mapping on the data at each time location in the core cache queue of the component region to generate a component region embedding sequence; Perform channel normalization and time dimension aggregation on the component region embedding sequence, and perform dimension transformation on the aggregation result to generate the core representation of the component region; The inverter cluster core representation and the site core representation are generated as follows: Read the inverter cluster number, module area number, inverter number and site number from the inverter cluster hierarchy data, and determine the core representation of the module area corresponding to the same inverter cluster according to the inverter cluster number; According to the component area numbering order, the core representations of multiple component areas corresponding to the same inverter cluster are arranged into an intra-cluster embedding matrix; Align the core representation of the component region in the cluster embedding matrix according to the acquisition time position; Perform channel normalization and dimension transformation on the aligned intra-cluster embedding matrix to generate the inverter cluster core representation; The core representations of multiple inverter clusters corresponding to the same site are determined according to the site number, and the core representations of multiple inverter clusters are arranged into a site embedding matrix according to the order of the inverter cluster numbers. Perform channel normalization and dimension transformation on the site embedding matrix to generate the site core representation; The component thermal inertia memory slot layer reads the component's thermal state data, establishes component temperature rise memory slots, component temperature drop memory slots, irradiation recovery hysteresis memory slots, and temperature rise power decay memory slots, and generates thermal inertia state embeddings according to the time sequence of each memory slot, wherein: The component's thermal inertia memory slot layer includes: Component temperature register: stores component area number, acquisition time, and component temperature value; Ambient temperature cache table: stores ambient temperature values, irradiance values, and corresponding acquisition times; Power response cache queue: Stores historical power generation values and power changes according to component region number; Component temperature rise memory slot: Stores temperature change records corresponding to the continuous rise of component temperature; Component cooling memory slot: Stores temperature change records corresponding to the continuous temperature drop phase of the component; Irradiation recovery hysteresis memory slot: stores the time location and power change records of the power response hysteresis after irradiance recovery; Temperature rise power decay memory slot: Records of power decrease and insufficient power recovery during the temperature rise phase of the storage component; Thermal state detector: Determines the thermal state type based on changes in component temperature, irradiance, and power. Thermal inertial embedding output: Integrates the thermal state records in each memory slot according to the acquisition time sequence to generate thermal inertial state embedding; In the component thermal inertia memory slot layer, the component temperature value in the component temperature register, the ambient temperature value and irradiance value in the ambient temperature cache table, and the historical power generation value and power change in the power response cache queue are input into the thermal state discriminator. The thermal state discriminator determines the component thermal state type based on the component temperature change, irradiance change, and power change, and writes the temperature change record corresponding to the heating stage into the heating memory slot, the temperature change record corresponding to the cooling stage into the cooling memory slot, the time position and power change record of the power response lag after irradiance recovery into the irradiance recovery lag memory slot, and the power decrease record corresponding to the component temperature rise stage into the temperature rise power decay memory slot. The thermal inertia embedding output reads the thermal state records in the heating memory slot, cooling memory slot, irradiance recovery lag memory slot, and temperature rise power decay memory slot in the order of acquisition time, and integrates them to generate thermal inertia state embedding. Establish component temperature rise memory slots and component temperature fall memory slots, specifically as follows: Read the component area number, acquisition time, and component temperature value from the component thermal status data, and arrange the component temperature values in order of component area number and acquisition time; Calculate the temperature change of the same component area between adjacent acquisition times; When the component temperature change is greater than 0, the corresponding component area number, current acquisition time, current component temperature value and component temperature change are written into the component temperature rise memory slot. When the component temperature change is less than 0, the corresponding component area number, current acquisition time, current component temperature value and component temperature change are written into the component cooling memory slot. When the component temperature change is equal to 0, the corresponding acquisition time position is marked as the temperature holding position, and the memory slot type corresponding to the previous acquisition time position is used. Organize the records in the component heating memory slot and component cooling memory slot according to the collection time sequence to generate the component heating memory sequence and component cooling memory sequence; Establish an irradiation recovery hysteresis memory slot, specifically as follows: Read the acquisition time, component area number, irradiance value, component temperature value and historical power generation value from the component thermal status data, and arrange them in order of component area number and acquisition time; Calculate the irradiance and power changes of the same component area between adjacent acquisition times, and mark the acquisition time locations where the irradiance change is greater than 80W / m² as irradiance recovery locations; At the sampling time corresponding to the irradiation recovery location, the power change of the corresponding component area is read. When the power change is less than or equal to 0, the corresponding sampling time location is marked as the first type of irradiation recovery lag location. The power variation was normalized according to the installed capacity of the station, and the irradiance variation was normalized according to 1000W / m². When the normalized power variation was less than 60% of the normalized irradiance variation, the corresponding acquisition time position was marked as the second type of irradiance recovery lag position. Write the component area number, acquisition time, irradiance value, component temperature value, historical power generation value, irradiance change and power change corresponding to the first type of irradiance recovery lag position and the second type of irradiance recovery lag position into the irradiance recovery lag memory slot. The records in the irradiation recovery lag memory slots are organized according to the acquisition time sequence to generate the irradiation recovery lag memory sequence; Establish a memory slot for temperature rise power decay, specifically as follows: Read the acquisition time, component area number, irradiance value, component temperature value, ambient temperature value, and historical power generation value from the component thermal status data, and arrange them in order of component area number and acquisition time; Calculate the component temperature change, ambient temperature change, irradiance change, and power change in the same component area between adjacent acquisition times; When the component temperature change is greater than 2℃ and the power change is less than 0, the corresponding sampling time location will be marked as the first type of temperature rise power decay location. When the component temperature change is greater than 2℃, and the irradiation change is greater than 0 and the power change is less than the power recovery reference value corresponding to the irradiation change, the corresponding sampling time position is marked as the second type of temperature rise power decay position. The power recovery reference value is determined based on the average correspondence between the irradiation change and the power change in the same component area within the first 6 sampling time positions. When the ambient temperature change is greater than 1℃, the component temperature change is greater than 2℃, and the power change is less than or equal to 0, the corresponding sampling time location will be marked as the third type of temperature rise power decay location. Write the component area number, acquisition time, irradiance value, component temperature value, ambient temperature value, historical power generation value, component temperature change and power change corresponding to the first type of temperature rise power decay location, the second type of temperature rise power decay location and the third type of temperature rise power decay location into the temperature rise power decay memory slot. Organize the records in the temperature rise power decay memory slot according to the acquisition time sequence to generate the temperature rise power decay memory sequence; Generate thermal inertial state embeddings, specifically as follows: Read the component heating memory sequence, component cooling memory sequence, irradiation recovery hysteresis memory sequence, and temperature rise power decay memory sequence, and arrange them according to the component area number and the order of acquisition time; Write the component temperature change, current component temperature value, and acquisition time position from the component temperature rise memory sequence into the temperature rise state vector; Write the component temperature change, current component temperature value, and acquisition time position from the component cooling memory sequence into the cooling status vector; Write the irradiation change, power change, component temperature value, and irradiation recovery hysteresis position from the irradiation recovery hysteresis memory sequence into the irradiation recovery hysteresis state vector. Write the component temperature change, ambient temperature change, power change, and temperature rise power decay position in the temperature rise power decay memory sequence into the temperature rise power decay state vector. The heating state vector, cooling state vector, irradiation recovery hysteresis state vector, and temperature rise power decay state vector are spliced together according to the acquisition time sequence to generate the component thermal inertia state sequence. Perform channel normalization and dimension transformation on the component thermal inertial state sequence to generate thermal inertial state embedding; The improved SOFTS network was trained using a combination of power prediction error, state switching core representation consistency error, inverter cluster hierarchical folding error, and thermal inertia state embedding error as joint optimization objectives. The network parameters of the photovoltaic state anchor core refresh layer, inverter cluster hierarchical core folding layer, and module thermal inertia memory slot layer were continuously optimized. When the change in the joint loss value over five consecutive training rounds was less than 0.001, the improved SOFTS network was considered to have completed convergence training. The improved SOFTS network is trained as follows: The photovoltaic state fusion feature set, actual power generation sequence, actual inverter cluster power sequence, photovoltaic state switching candidate points, and module thermal state records are read from the training samples. The photovoltaic state fusion feature set is input into the improved SOFTS network. After passing through the photovoltaic state anchor core refresh layer, the state switching core representation is output. After passing through the inverter cluster level core folding layer, the station core representation is output. After passing through the module thermal inertia memory slot layer, the thermal inertia state embedding is output. The station power prediction sequence and inverter cluster power prediction sequence corresponding to the next 6 prediction time steps are generated through the prediction output terminal. During the training process, each input window contains 24 consecutive acquisition time positions, and 64 input windows are read in each batch. The power prediction error is obtained by averaging the squared difference between the power prediction sequence and the actual power generation sequence. The consistency error of the state switching core representation is obtained by averaging the squared difference between the prediction anchor point position corresponding to the state switching core representation and the acquisition time position corresponding to the photovoltaic state switching candidate point. The average value is obtained by squaring the difference between the inverter cluster power prediction sequence and the actual inverter cluster power sequence, and by squaring the difference between the sum of the power prediction values of each inverter cluster under the same site and the site power prediction value, to obtain the inverter cluster hierarchical folding error. The thermal inertia state embedding error is obtained by averaging the squared differences between the corresponding temperature rise power attenuation mark, irradiation recovery hysteresis mark and the corresponding mark in the component thermal state record. The power prediction error is multiplied by 0.45, the state switching core representation consistency error by 0.20, the inverter cluster hierarchical folding error by 0.20, and the thermal inertia state embedding error by 0.15, and then summed to generate a joint loss value. The loss change rate corresponding to each network parameter in the photovoltaic state anchor core refresh layer, inverter cluster hierarchical core folding layer, and module thermal inertia memory slot layer is read from the joint loss value to obtain the gradient value corresponding to each network parameter. The gradient value is multiplied by 0.001 and then subtracted from the current network parameter value to obtain the updated network parameter value, which is then rewritten into the improved SOFTS network. When the change amplitude of the joint loss value in 5 consecutive training rounds is less than 0.001, the improved SOFTS network is considered to have completed convergence training.
[0025] In this embodiment, generating the solar power generation prediction result includes: Read the state switching core representation, the station core representation, the thermal inertia state embedding, and the photovoltaic state switching candidate points, and generate a prediction time step identifier according to the prediction task, and generate a prediction input representation sequence according to the prediction time step identifier; Map the state switching core representation to the state switching prediction vector, map the station core representation to the station power prediction vector, and map the thermal inertial state embedding to the thermal inertial correction vector. Based on the predicted time step order, the state transition prediction vector, the station power prediction vector, and the thermal inertia correction vector are mapped to generate a predicted time step power representation sequence, where: Generate the predicted time step power representation sequence, specifically: Read the prediction time step identifiers and build a prediction time step index table according to the prediction time step order; The state switching prediction vector, the station power prediction vector, and the thermal inertia correction vector are matched according to the prediction time step index table to generate a prediction time step vector group. For each prediction time step vector group, the state switching prediction vector, the station power prediction vector, and the thermal inertia correction vector are concatenated in the order of vector dimensions to generate the prediction time step fusion vector. Perform channel normalization and dimension transformation on the predicted time step fusion vector to generate the predicted time step power representation; Arrange all prediction time step power representations in the order of prediction time steps to generate a sequence of prediction time step power representations; A power prediction sequence for the power plant is generated based on the predicted time step power representation sequence, and a power prediction sequence for the inverter cluster is generated based on the power plant equipment connection relationship table, wherein: The power station power prediction sequence is generated as follows: Read the power representation sequence of the prediction time steps and extract the power representation of each prediction time step according to the prediction time step order; perform dimensional transformation on the power representation of each prediction time step to generate the predicted power value of the corresponding prediction time step; arrange the predicted power values of each prediction time step according to the prediction time step order to generate the prediction sequence of the power station; associate the prediction sequence of the power station with the prediction time step identifier to form the prediction sequence of the power station. The inverter cluster power prediction sequence is generated as follows: Read the power prediction sequence and the equipment connection table of the power plant, and determine the power plant number, inverter number, component area number and installed capacity data according to the equipment connection table; The component areas connected to the same inverter number are grouped into the same inverter cluster, and the installed capacity of each component area within the same inverter cluster is added together to obtain the installed capacity of the inverter cluster. The installed capacity of the station is obtained by adding up the installed capacity of each inverter cluster within the same station. Divide the installed capacity of the inverter cluster by the installed capacity of the power station to obtain the capacity ratio of the inverter cluster; Read the predicted power values of the power plants in the power prediction sequence according to the prediction time step order, and multiply the predicted power value of the power plants in each prediction time step by the inverter cluster capacity ratio to obtain the initial cluster power prediction value of the corresponding prediction time step. Arrange the initial cluster power prediction values corresponding to the same inverter cluster in the order of prediction time steps to generate a single inverter cluster power prediction sequence. All single inverter cluster power prediction sequences are compiled according to inverter numbering order to generate inverter cluster power prediction sequences; The power prediction sequence of the power plant, the power prediction sequence of the inverter cluster, the prediction time step identifier, and the photovoltaic state switching candidate points are combined to generate the solar power generation prediction result.
[0026] In this embodiment, the step of collecting and updating actual power generation feedback data includes: Collect actual power generation feedback data, actual meteorological feedback data, component temperature feedback data, and inverter status feedback data within the forecast period; Based on the site number, inverter number, component area number, acquisition time, and prediction time step identifier, the actual power generation feedback data and the solar power generation prediction results are matched to generate a prediction result correspondence table. Based on the prediction result correspondence table, the difference between the actual power generation and the corresponding predicted power value is calculated, a prediction error sequence is generated, and the prediction error sequence, actual meteorological feedback data, component temperature feedback data and inverter status feedback data are combined to generate a prediction feedback record. The standardized photovoltaic prediction dataset is updated based on the prediction feedback records, and the convolution kernel set parameters, pooling statistics parameters, and response permutation encoding parameters of the MultiROCKET algorithm are updated based on the prediction error sequence. Based on the prediction feedback records, update and improve the parameters of the photovoltaic state anchor core refresh layer, inverter cluster level core fold layer, and module thermal inertia memory slot layer of the SOFTS network.
[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous photovoltaic power prediction operation cycle. Power plant A was connected to the method of this invention for short-term power generation prediction. Power plant A has an installed capacity of 20MW, comprising 4 inverter clusters and 16 module areas, with each inverter cluster connected to 4 module areas. The sampling interval was 5 minutes. Each training sample read 24 past sampling points and predicted 6 future sampling points, corresponding to a power prediction for the next 30 minutes. The training sample contained 17,280 sampling records, the validation sample contained 2,880 sampling records, and the test sample contained 2,880 sampling records. The original data contained short-term missing irradiance data, delayed module temperature reporting, and abnormal power jumps. The missing irradiance data rate was 1.7%, the missing module temperature data rate was 1.1%, and the percentage of abnormal power jump records was 0.6%.
[0028] During data preprocessing, this method correlates multi-source photovoltaic (PV) prediction data according to acquisition time, site number, inverter number, module region number, and data type identifier. A certain module region lacks an irradiance value at sampling time T08; it is 764 W / m² at T07 and 721 W / m² at T09. After interpolation, T08 shows 742.5 W / m². Inverter cluster I02's power record at T15 is 5.92 MW, while it is 3.41 MW at T14 and 3.36 MW at T16. This record is identified as an abnormal jump value and corrected to 3.38 MW. After completing duplicate record deletion, invalid record removal, missing value imputation, timestamp alignment, and numerical standardization, the data integrity rate increases from 96.6% to 99.7%, and the timestamp alignment rate increases from 94.8% to 100%, forming a standardized PV prediction dataset.
[0029] A multi-source photovoltaic time series was constructed based on a standardized photovoltaic prediction dataset. In a certain training sample, the power series from T01 to T24 changed from 13.82MW to 9.74MW, the irradiance series changed from 795W / m² to 436W / m², the cloud cover increased from 22% to 73%, and the module temperature increased from 41.6℃ to 45.2℃ and then decreased to 44.1℃. Simultaneously, the processing flow read the connection relationships between module areas, inverters, and the power plant, obtaining that C01 to C04 are connected to I01, C05 to C08 are connected to I02, C09 to C12 are connected to I03, and C13 to C16 are connected to I04. The four inverter clusters converge at power plant A, thereby generating a power plant equipment connection relationship table.
[0030] After inputting the multi-source photovoltaic time series data into the MultiROCKET algorithm, this method performs a convolutional transformation of the irradiance-power phase diagram trajectory. Taking T05 to T09 as an example, the irradiance decreases from 782 W / m² to 418 W / m², and the power decreases from 14.66 MW to 9.28 MW; in the T10 to T14 stage, the irradiance recovers from 452 W / m² to 816 W / m², and the power recovers from 9.91 MW to 14.51 MW. The processing combines the irradiance and power changes at adjacent sampling times into irradiance-power phase diagram points, forming a trajectory sequence of 24 phase diagram points, which is then converted into a two-channel trajectory sequence. After performing convolutional response calculation on the two-channel trajectory sequence using a set of convolutional kernels, in the continuous decreasing segment from T06 to T09, the positive proportional pooling result is 0.78, and the longest continuous positive pooling result is 4 sampling points, finally generating an irradiance-power response trajectory feature matrix with a dimension of 24×256.
[0031] In the cross-variable homo-kernel response permutation coding process, the same set of convolutional kernels are applied to the power time series, irradiance time series, cloud cover time series, module temperature time series, wind speed and direction time series, and inverter state time series, respectively. One convolutional kernel has a response position of T05 in the cloud cover time series, T06 in the irradiance time series, and T07 in the power time series, forming a cloud cover-irradiance-power response permutation. Another convolutional kernel shows a recovery response at T11 in the irradiance time series, a sustained response from T12 to T14 in the module temperature time series, and a insufficient recovery response at T13 in the power time series, forming an irradiance-module temperature-power response permutation. After processing, the cross-variable homo-kernel response permutation coding table contains 960 coded records and generates photovoltaic state switching candidate points located near T06, T09, and T13.
[0032] Based on the photovoltaic state transition candidate points, this method performs anchor point segmentation on the multi-source photovoltaic time series. Using T06 as the anchor point, T01 to T05 are divided into the pre-anchor point sequence, T07 to T11 into the post-anchor point sequence, and T04 to T08 into the cross-anchor point sequence. The processing associates the anchor point segmented sequences with the corresponding module region number, inverter number, and site number, and writes the irradiance-power response trajectory feature matrix and the cross-variable co-core response permutation encoding table to the corresponding acquisition time position, forming the anchor point state feature table. In the T10 to T14 stage, the module temperature remains between 44.1℃ and 46.2℃, and the irradiance has recovered to above 800W / m². This segment is written into the module thermal state data, ultimately generating the photovoltaic state fusion feature set.
[0033] After the photovoltaic state fusion feature set is input into the improved SOFTS network, the photovoltaic state anchor core refresh layer generates a state switching core representation based on the anchor segment sequence and photovoltaic state switching candidate points. At T06, the cloud shadow entry response intensity in the cross-anchor core representation is 0.76, and the stable power generation response intensity is 0.22. The inverter cluster-level core folding layer generates a site core representation according to the hierarchical relationship from the module region to the inverter and from the inverter to the site. In the test sample, the I03 inverter cluster is significantly affected by local cloud shadows, and the actual average power drops from 3.11MW to 1.92MW. The site core representation retains the local decrease response of I03. The module thermal inertia memory slot layer reads an irradiance of 816W / m², a module temperature of 46.2℃, and an actual power of 14.51MW at T13. The state of irradiance recovery but insufficient power recovery is written into the temperature rise power decay memory slot to generate thermal inertia state embedding.
[0034] In the prediction phase, this method generates predicted solar power generation results for six future sampling points. The actual power of this test sample is 14.72MW, 14.95MW, 15.08MW, 15.21MW, 15.18MW, and 15.06MW, respectively. The predicted values of traditional multivariate prediction methods are 15.61MW, 15.88MW, 16.04MW, 16.12MW, 16.03MW, and 15.92MW, respectively; the predicted values of ordinary SOFTS networks are 15.24MW, 15.39MW, 15.51MW, 15.56MW, 15.49MW, and 15.42MW, respectively; and the predicted values of the method of this invention are 14.86MW, 15.02MW, 15.16MW, 15.28MW, 15.22MW, and 15.11MW, respectively. In this sample, the average absolute error of the traditional method over 6 steps was 0.91MW, that of the ordinary SOFTS network was 0.37MW, and that of the method of this invention was 0.08MW.
[0035] In the overall comparison of the test samples, the traditional multivariate prediction method had a mean absolute error of 0.82MW, a root mean square error of 1.10MW, and a mean absolute percentage error of 7.8% for power generation; the ordinary SOFTS network had a mean absolute error of 0.61MW, a root mean square error of 0.84MW, and a mean absolute percentage error of 5.6% for power generation; and the method of this invention had a mean absolute error of 0.38MW, a root mean square error of 0.52MW, and a mean absolute percentage error of 3.4% for power generation. In the cloud shadow entry sample, the traditional method had a mean error of 1.43MW, the ordinary SOFTS network had a mean error of 1.02MW, and the method of this invention had a mean error of 0.59MW. In the cloud cover recovery sample, the traditional method had a mean error of 1.28MW, the ordinary SOFTS network had a mean error of 0.91MW, and the method of this invention had a mean error of 0.55MW. Regarding the accuracy of power ramp direction identification, the traditional method achieved 82.6%, the ordinary SOFTS network achieved 88.4%, and the method of this invention achieved 94.8%.
[0036] After the prediction is completed, this method collects actual power generation feedback data, actual meteorological feedback data, module temperature feedback data, and inverter status feedback data to generate a prediction result correspondence table. The processing calculates the difference between the actual power generation and the corresponding predicted power value, forming a prediction error sequence, which is then written into the prediction feedback record. After five consecutive rounds of feedback updates, 1440 new feedback records are added to the standardized photovoltaic prediction dataset, and the MultiROCKET algorithm parameters and improved SOFTS network parameters are updated. Before the feedback update, the average absolute error of the method on cloud shadow disturbance samples was 0.59MW, which decreased to 0.51MW after the feedback update; the prediction lag near state switching decreased from 1.1 sampling points to 0.6 sampling points. As can be seen from Example 1, the method of this invention has a clear data change process in each processing step, and can effectively handle cloud shadow disturbances, photovoltaic state switching, local fluctuations in inverter clusters, and the effects of module thermal inertia, verifying the engineering feasibility and prediction effectiveness of this method in solar power generation prediction scenarios.
[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for predicting solar power generation capacity based on big data analysis, characterized in that, include: Multi-source photovoltaic (PV) prediction data from solar power plants is collected and preprocessed to generate a standardized PV prediction dataset. Based on this dataset, a multi-source PV time series is constructed, and a connection table for plant equipment is generated according to the connection relationships between component regions, inverters, and the plant. The multi-source PV time series is input into the MultiROCKET algorithm to perform irradiance-power phase diagram trajectory convolution transformation and cross-variable co-core response permutation coding, generating an irradiance-power response trajectory feature matrix, a cross-variable co-core response permutation coding table, and PV state switching candidate points. Anchor point segmentation is performed on the multi-source PV time series based on these candidate points, and a PV state fusion feature set is generated by combining the plant equipment connection table, the irradiance-power response trajectory feature matrix, and the cross-variable co-core response permutation coding table. An improved SOWTS network, comprising a PV state anchor point core refresh layer, an inverter cluster-level core folding layer, and a component thermal inertia memory slot layer, is constructed. This network is then input into the PV state fusion feature set to generate a state switching core representation, a plant core representation, and a thermal inertia state embedding. Based on the state switching core representation, the station core representation, and the thermal inertia state embedding, the solar power generation prediction results are generated; actual power generation feedback data are collected, prediction feedback records are generated, and the standardized photovoltaic prediction dataset, MultiROCKET algorithm parameters, and improved SOFTS network parameters are updated according to the prediction feedback records.
2. The solar power generation prediction method based on big data analysis according to claim 1, characterized in that, The generation of the standardized photovoltaic prediction dataset includes: Collect multi-source photovoltaic forecast data from solar power plants. The multi-source photovoltaic forecast data includes historical power generation data, irradiance data, cloud cover data, ambient temperature data, module temperature data, wind speed and direction data, humidity data, inverter status data, module regional operation data, plant installed capacity data, and weather forecast data. The multi-source photovoltaic prediction data are correlated according to the collection time, site number, inverter number, component area number and data type identifier to generate a photovoltaic raw prediction data sequence. The original photovoltaic prediction data sequence is processed by deleting duplicate records, removing invalid records, handling abnormal jump values, and filling missing values to generate a cleaned photovoltaic prediction data sequence. The cleaned photovoltaic prediction data sequence is processed by timestamp alignment, sampling interval unification, unit unification and numerical standardization to generate a standardized photovoltaic prediction data sequence. The standardized photovoltaic forecast data sequence is archived according to the site number, inverter number, module area number and collection time to generate a standardized photovoltaic forecast dataset.
3. The solar power generation prediction method based on big data analysis according to claim 1, characterized in that, The construction of the multi-source photovoltaic time series and the generation of the power station equipment connection table include: Read the site number, inverter number, module area number, acquisition time and data type identifier from the standardized photovoltaic prediction dataset, and generate a photovoltaic data index table; According to the photovoltaic data index table, historical power generation data, irradiance data, cloud cover data, ambient temperature data, module temperature data, wind speed and direction data, humidity data, and inverter status data are written into the corresponding time positions to generate a multi-source photovoltaic time series. The multi-source photovoltaic time series were arranged according to the collection time sequence, and then divided into power time series, irradiance time series, cloud cover time series, ambient temperature time series, module temperature time series, wind speed and direction time series, humidity time series, and inverter status time series according to the data type identifier. Read the connection relationship data between the module area, inverter and the station in the solar power plant, and establish the correspondence between the module area number and the inverter number, and the correspondence between the inverter number and the station number; The component area number, inverter number, site number, installed capacity data, and connection level identifier are associated to generate a site equipment connection relationship table.
4. The solar power generation prediction method based on big data analysis according to claim 1, characterized in that, The generation of the irradiance-power response trajectory feature matrix, the transvariable homogeneous response permutation coding table, and photovoltaic state switching candidate points includes: Read the power time series, irradiance time series, cloud cover time series, module temperature time series, wind speed and direction time series, and inverter status time series from the multi-source photovoltaic time series, and construct the MultiROCKET input window according to the acquisition time order; The power change and irradiance change at adjacent acquisition times are calculated for the power time series and irradiance time series respectively. The irradiance change and power change at the same acquisition time are combined into irradiance-power phase diagram points, and the irradiance-power phase diagram trajectory sequence is generated according to the acquisition time sequence. The irradiation-power phase diagram trajectory sequence is subjected to irradiation-power phase diagram trajectory convolution transformation, that is, the irradiation-power phase diagram trajectory sequence is converted into a two-channel trajectory sequence. The convolution response is calculated on the two-channel trajectory sequence through the convolution kernel set in the MultiROCKET algorithm, and the irradiation-power response trajectory feature matrix is generated by positive proportional pooling, positive mean pooling, positive index mean pooling and longest continuous positive pooling. Perform cross-variable homogeneous kernel response permutation coding on power time series, irradiance time series, cloud cover time series, component temperature time series, wind speed and direction time series and inverter state time series. Apply the same set of convolution kernels to each time series and record the response position and response order of the same convolution kernel in different time series. The cross-variable response results of the same convolution kernel are arranged according to the convolution kernel number, time series type, response position and response order to generate a cross-variable same kernel response arrangement coding table; Based on the convolution response sign switching position in the irradiance-power response trajectory feature matrix, the boundary of the longest continuous positive value interval, and the response arrangement change position in the cross-variable homogeneous response arrangement coding table, photovoltaic state switching candidate points are generated according to the acquisition time.
5. The solar power generation prediction method based on big data analysis according to claim 1, characterized in that, The generated photovoltaic state fusion feature set includes: Read the multi-source photovoltaic time series, photovoltaic state switching candidate points, station equipment connection relationship table, irradiance-power response trajectory feature matrix and cross-variable homogeneous response permutation coding table; The photovoltaic state switching candidate points are sorted according to the acquisition time sequence, and the photovoltaic state switching candidate points are mapped to the acquisition time positions in the multi-source photovoltaic time series to generate a photovoltaic state anchor point index table. Anchor point segmentation is performed on the multi-source photovoltaic time series according to the photovoltaic state anchor point index table to generate an anchor point segmented sequence including the pre-anchor point sequence, the post-anchor point sequence, and the cross-anchor point sequence; Based on the site equipment connection relationship table, the anchor point segment sequence is associated with the corresponding component area number, inverter number and site number respectively, and an anchor point segment equipment association table is generated. The irradiation-power response trajectory feature matrix and the transvariable homonuclear response arrangement coding table are written into the anchor point segmentation equipment association table according to the acquisition time and position to generate the anchor point status feature table. Based on the anchor point state characteristic table, photovoltaic state anchor point index table, site equipment connection relationship table, and component temperature time series in multi-source photovoltaic time series, a photovoltaic state fusion feature set is generated, which includes anchor point segment sequence, photovoltaic state switching candidate points, inverter cluster hierarchical relationship data, component thermal state data, irradiance-power response trajectory characteristics, and cross-variable co-core response permutation coding.
6. The solar power generation prediction method based on big data analysis according to claim 1, characterized in that, The generation of the state switching core representation, the station core representation, and the thermal inertial state embedding includes: An improved SOFTS network is constructed, which includes a photovoltaic state anchor core refresh layer, an inverter cluster-level core folding layer, and a module thermal inertia memory slot layer. The photovoltaic state anchor core refresh layer reads the anchor segment sequence and photovoltaic state switching candidate points, generates the core representation before the anchor point, the core representation after the anchor point, and the core representation across anchor points, respectively, and refreshes the core representation across anchor points according to the acquisition time position corresponding to the photovoltaic state switching candidate point to generate the state switching core representation; The inverter cluster hierarchical core folding layer reads the inverter cluster hierarchical relationship data and generates the component area core representation, inverter cluster core representation and site core representation according to the hierarchical relationship from the component area to the inverter and from the inverter to the site. The component thermal inertia memory slot layer reads the component thermal state data, establishes component heating memory slots, component cooling memory slots, irradiation recovery hysteresis memory slots, and temperature rise power decay memory slots, and generates thermal inertia state embeddings according to the time sequence of each memory slot. The improved SOFTS network was trained using a combination of power prediction error, state switching core representation consistency error, inverter cluster hierarchical folding error, and thermal inertia state embedding error as joint optimization objectives. The network parameters of the photovoltaic state anchor core refresh layer, inverter cluster hierarchical core folding layer, and module thermal inertia memory slot layer were continuously optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved SOFTS network was considered to have completed convergence training.
7. The solar power generation prediction method based on big data analysis according to claim 1, characterized in that, The generated solar power generation prediction results include: Read the state switching core representation, the station core representation, the thermal inertia state embedding, and the photovoltaic state switching candidate points, and generate a prediction time step identifier according to the prediction task, and generate a prediction input representation sequence according to the prediction time step identifier; Map the state switching core representation to the state switching prediction vector, map the station core representation to the station power prediction vector, and map the thermal inertial state embedding to the thermal inertial correction vector. According to the prediction time step order, the state switching prediction vector, the station power prediction vector, and the thermal inertia correction vector are matched to generate the prediction time step power representation sequence; The power prediction sequence for the power plant is generated based on the predicted time step power representation sequence, and the power prediction sequence for the inverter cluster is generated based on the power plant equipment connection relationship table. The power prediction sequence of the power plant, the power prediction sequence of the inverter cluster, the prediction time step identifier, and the photovoltaic state switching candidate points are combined to generate the solar power generation prediction result.
8. The solar power generation prediction method based on big data analysis according to claim 1, characterized in that, The process of collecting and updating actual power generation feedback data includes: Collect actual power generation feedback data, actual meteorological feedback data, component temperature feedback data, and inverter status feedback data within the forecast period; Based on the site number, inverter number, component area number, acquisition time, and prediction time step identifier, the actual power generation feedback data and the solar power generation prediction results are matched to generate a prediction result correspondence table. Based on the prediction result correspondence table, the difference between the actual power generation and the corresponding predicted power value is calculated, a prediction error sequence is generated, and the prediction error sequence, actual meteorological feedback data, component temperature feedback data and inverter status feedback data are combined to generate a prediction feedback record. The standardized photovoltaic prediction dataset is updated based on the prediction feedback records, and the convolution kernel set parameters, pooling statistics parameters, and response permutation encoding parameters of the MultiROCKET algorithm are updated based on the prediction error sequence. Based on the prediction feedback records, update and improve the parameters of the photovoltaic state anchor core refresh layer, inverter cluster level core fold layer, and module thermal inertia memory slot layer of the SOFTS network.
