Satellite thermal infrared cloud event fusion-based power prediction method and system for hydropower station

Abstract

The embodiment of the application discloses a kind of power station power prediction method and system based on satellite thermal infrared cloud event fusion, the method includes the following steps: S1, obtain the historical active power sequence before prediction starting point and satellite thermal infrared brightness temperature image sequence;S2, historical active power sequence is encoded into hydropower operation state segment token;S3, by the latest power statistics generation hydropower state query vector;S4, cold cloud coverage, cold cloud depth, space texture and time evolution feature are extracted from brightness temperature image to form satellite cloud event token;S5, using hydropower state query vector obtains state-guided satellite context;S6, by each hydropower operation state segment token carries out segment-level cross fusion to satellite token;S7, according to the encoding result output the daily average active power prediction value of the future natural day.The application can fuse cloud system evolution information and power station current operation state in prediction, and can reduce active power prediction error.

Description

Technical Field

[0001] This invention relates to the fields of active power prediction for hydropower stations, satellite remote sensing data processing, and smart grid dispatching, and particularly to a method and system for predicting hydropower station power based on satellite thermal infrared cloud event fusion. Background Technology

[0002] Short-term active power forecasting for hydropower stations is used for power system dispatching, reserve capacity allocation, renewable energy coordination, and hydropower resource operation and management. Existing forecasting methods based on historical power series can utilize the recent operating status of the power station, but during the development of strong cloud systems, cold cloud events, and precipitation-related hydrological responses, relying solely on historical power series tends to lag behind upstream weather processes.

[0003] On the other hand, traditional meteorological inputs are often compressed into point observations, scalar forecasts, or grid statistics, making it difficult to retain information on cold cloud top coverage, cloud core depth, spatial texture, and temporal evolution of cloud systems in thermal infrared cloud images. Therefore, a forecasting method is needed that can deeply integrate cloud event evidence from satellite thermal infrared brightness temperature images with the current operating status of power plants to output daily average active power forecasts for future natural days, thereby reducing active power forecasting errors. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide a method and system for predicting the power of hydropower stations based on satellite thermal infrared cloud event fusion, so as to output the daily average active power prediction value for future natural days and reduce the active power prediction error.

[0005] To address the aforementioned technical problems, this invention proposes a hydropower station power prediction method based on satellite thermal infrared cloud event fusion, comprising the following steps: S1. Obtain the historical active power sequence before the prediction start point and the satellite thermal infrared brightness temperature image sequence within the corresponding time range, and determine the future natural days to be predicted; S2. The historical active power sequence is normalized and divided into overlapping hour segments, and each overlapping hour segment is mapped to a hydropower operation status segment token containing a location code. S3. Construct a hydropower operation status descriptor based on the nearest window of the historical active power sequence, and map the hydropower operation status descriptor into a hydropower status query vector; S4. Perform effective pixel processing on the satellite thermal infrared brightness temperature image sequence, and extract cold cloud coverage, cold cloud depth, spatial texture, regional statistics and temporal evolution features from the brightness temperature image to form a time-aligned satellite cloud event token sequence. S5. Perform time attention on the satellite cloud event token sequence using the hydropower status query vector to obtain the satellite context guided by hydropower status, and merge the satellite context guided by hydropower status with the satellite cloud event tokens at each time moment into a hydropower status conditional satellite token. S6. Using each hydropower operation status segment token as a query, perform segment-level cross-attention on the hydropower status conditional satellite token to obtain the satellite context corresponding to each hydropower operation status segment token, merge the corresponding hydropower operation status segment token and satellite context into a fused segment token, and encode the fused segment token. S7. Output the predicted daily average active power for the future natural days based on the encoding results.

[0006] Accordingly, embodiments of the present invention also provide a hydropower station power prediction system based on satellite thermal infrared cloud event fusion, comprising: Data acquisition module: acquires the historical active power sequence before the prediction start point and the satellite thermal infrared brightness temperature image sequence within the corresponding time range, and determines the future natural days to be predicted; Hydropower Segment Token Module: Normalizes the historical active power sequence into instances and divides it into overlapping hour segments, mapping each overlapping hour segment into a hydropower operation status segment token containing a location code; Hydropower status query module: Constructs a hydropower operation status descriptor based on the most recent window of the historical active power sequence, and maps the hydropower operation status descriptor to a hydropower status query vector; Satellite cloud event token module: Performs effective pixel processing on the satellite thermal infrared brightness temperature image sequence, and extracts cold cloud coverage, cold cloud depth, spatial texture, regional statistics and temporal evolution features from the brightness temperature image to form a time-aligned satellite cloud event token sequence; Hydropower guidance context module: Utilizes the hydropower status query vector to perform time attention on the satellite cloud event token sequence to obtain the hydropower status-guided satellite context, and merges the hydropower status-guided satellite context with the satellite cloud event tokens at each time moment into a hydropower status-conditionalized satellite token; Fragment-level cross-fusion module: Using each hydropower operation status fragment token as a query, perform fragment-level cross-attention on the hydropower status conditional satellite token to obtain the satellite context corresponding to each hydropower operation status fragment token, fuse the corresponding hydropower operation status fragment token and satellite context into a fused fragment token, and encode the fused fragment token; Power prediction module: Outputs the predicted daily average active power for the next natural day based on the encoding results.

[0007] The beneficial effects of this invention are as follows: Compared to methods that rely solely on historical power sequences or use meteorological information as a scalar input, this invention preserves physical cloud event information in satellite thermal infrared cloud images by utilizing cold cloud coverage, cold cloud depth, spatial texture, and temporal evolution characteristics.

[0008] This invention uses hydropower status query vectors to perform state-related selection on satellite cloud event time series, so that the same satellite cloud event has different context weights under different power plant operating states.

[0009] This invention employs fragment-level cross-fusion to select corresponding satellite cloud event evidence for each historical active power fragment, thereby jointly modeling historical power patterns and cloud system evolution information within a single direct prediction model. Fixed-month hold-out experiments and event-stratified diagnostics in embodiments of this invention demonstrate that this fusion path can reduce active power prediction errors in cloud event-related samples and medium- to long-term prediction scenarios. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the hydropower prediction method based on satellite thermal infrared cloud event fusion according to an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of the structure of a hydropower station power prediction system based on satellite thermal infrared cloud event fusion according to an embodiment of the present invention. Detailed Implementation

[0012] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] In this embodiment of the invention, directional indicators (such as up, down, left, right, front, back, etc.) are only used to explain the relative positional relationship and movement of each component in a specific posture (as shown in the figure). If the specific posture changes, the directional indicator will also change accordingly.

[0014] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0015] Please refer to Figure 1 The hydropower prediction method based on satellite thermal infrared cloud event fusion in this embodiment of the invention includes steps S1 to S7.

[0016] S1. Obtain the historical active power sequence before the prediction start point and the satellite thermal infrared brightness temperature image sequence within the corresponding time range, and determine the future natural day to be predicted.

[0017] S2. The historical active power sequence is instance-normalized and divided into overlapping hour segments, and each overlapping hour segment is mapped to a hydropower operation status segment token containing a location code. As one implementation, the historical active power sequence includes active power records of the power plant for several consecutive hours prior to the prediction start point, the satellite thermal infrared brightness temperature image sequence includes multi-channel thermal infrared brightness temperature image blocks prior to the prediction start point, and the predicted daily average active power value is the average of the active power records for multiple hours within the future natural day.

[0018] In one implementation, step S2 involves instance normalization based on the mean and variance of the historical active power sequence; forming multiple overlapping hour segments according to a preset segment length and a preset step size; performing linear mapping on each overlapping hour segment and superimposing learnable position embeddings to obtain the hydropower operation status segment token.

[0019] S3. Construct a hydropower operation status descriptor based on the most recent window of the historical active power sequence, and map the hydropower operation status descriptor into a hydropower status query vector.

[0020] In one implementation, the hydropower operation status descriptor includes the mean of the most recent window, the standard deviation of the most recent window, the active power at the prediction starting point, the difference between the active power at the beginning and end of the most recent window, the difference between the mean of the most recent window and the mean of the previous reference window, and the average absolute ramp of the most recent window; wherein, the most recent window is a sample of the 24 hours before the prediction starting point, and the previous reference window is a sample of the 24 hours immediately preceding the most recent window.

[0021] S4. Perform effective pixel processing on the satellite thermal infrared brightness temperature image sequence, and extract cold cloud coverage, cold cloud depth, spatial texture, regional statistics and temporal evolution features from the brightness temperature image to form a time-aligned satellite cloud event token sequence.

[0022] In one implementation, step S4 involves identifying the effective pixels of each satellite thermal infrared brightness temperature image; within a preset spatial partition, determining the cold cloud coverage based on the proportion of effective pixels with brightness temperatures below the cold cloud threshold, and determining the cold cloud depth based on the positive difference between the cold cloud brightness temperature threshold and the effective pixel brightness temperature; and extracting at least one cloud event descriptor from the effective pixel proportion, local spatial texture, global channel statistics, regional mean, and cold anomaly intensity.

[0023] The time evolution features include the first-order difference of the satellite cloud event descriptor, the deviation relative to the retrospective mean, the short-window smoothing result, the long-window smoothing result, and the cumulative mean result, and the time evolution features are interpolated to the same time length as the historical active power sequence.

[0024] S5. Perform time attention on the satellite cloud event token sequence using the hydropower status query vector to obtain the satellite context guided by the hydropower status, and merge the satellite context guided by the hydropower status with the satellite cloud event tokens at each time moment into a hydropower status conditional satellite token.

[0025] In one implementation, in step S5, the hydropower status query vector is mapped to a query vector, and the satellite cloud event token is mapped to a key vector and a value vector; the temporal attention weight is calculated based on the similarity between the query vector and each key vector; and the value vectors are weighted and summed based on the temporal attention weight to obtain the satellite context guided by the hydropower status.

[0026] S6. Using each hydropower operation status segment token as a query, perform segment-level cross-attention on the hydropower status conditional satellite token to obtain the satellite context corresponding to each hydropower operation status segment token, merge the corresponding hydropower operation status segment token and satellite context into a fused segment token, and encode the fused segment token.

[0027] In one implementation, in step S6, the hydropower state conditional satellite token is projected to the same hidden dimension as the hydropower operation state fragment token; for each hydropower operation state fragment token, the cross-attention weight between it and each hydropower state conditional satellite token is calculated; the satellite context corresponding to the hydropower operation state fragment token is obtained according to the cross-attention weight, and the satellite context is concatenated with the hydropower operation state fragment token and then fused and projected.

[0028] S7. The predicted daily average active power for the future natural days is output after encoding and prediction.

[0029] As one implementation method, after encoding the fused fragment token, a normalized prediction scalar is obtained through the prediction head, and the normalized prediction scalar is denormalized using the mean and standard deviation of the historical active power sequence to obtain the predicted value of the daily average active power for the future natural days; during training, the squared error between the predicted value and the actual daily average active power for the future natural days is used as the training loss.

[0030] Please refer to Figure 2 The hydropower station power prediction system based on satellite thermal infrared cloud event fusion according to an embodiment of the present invention includes: Data acquisition module: acquires the historical active power sequence before the prediction start point and the satellite thermal infrared brightness temperature image sequence within the corresponding time range, and determines the future natural days to be predicted; Hydropower Segment Token Module: Normalizes the historical active power sequence into instances and divides it into overlapping hour segments, mapping each overlapping hour segment into a hydropower operation status segment token containing a location code; Hydropower status query module: Constructs a hydropower operation status descriptor based on the most recent window of the historical active power sequence, and maps the hydropower operation status descriptor to a hydropower status query vector; Satellite cloud event token module: Performs effective pixel processing on the satellite thermal infrared brightness temperature image sequence, and extracts cold cloud coverage, cold cloud depth, spatial texture, regional statistics and temporal evolution features from the brightness temperature image to form a time-aligned satellite cloud event token sequence; Hydropower guidance context module: Utilizes the hydropower status query vector to perform time attention on the satellite cloud event token sequence to obtain the hydropower status-guided satellite context, and merges the hydropower status-guided satellite context with the satellite cloud event tokens at each time moment into a hydropower status-conditionalized satellite token; Fragment-level cross-fusion module: Using each hydropower operation status fragment token as a query, perform fragment-level cross-attention on the hydropower status conditional satellite token to obtain the satellite context corresponding to each hydropower operation status fragment token, fuse the corresponding hydropower operation status fragment token and satellite context into a fused fragment token, and encode the fused fragment token; Power prediction module: Outputs the predicted daily average active power for the next natural day based on the encoding results.

[0031] Example 1

[0032] This embodiment uses hourly active power records from the power plant side and FY-4B geostationary satellite thermal infrared brightness temperature images as inputs to illustrate an implementable process of the method of the present invention. The prediction start point is denoted as t. The historical active power sequence can be taken from the hourly active power records of the consecutive 7 days prior to the prediction start point. The satellite thermal infrared brightness temperature image sequence can be taken from satellite image blocks of the consecutive 7 days prior to the prediction start point. The satellite images can be sampled at three-hour intervals. The above window length and sampling interval are settings for this embodiment and do not limit the scope of protection of the present invention.

[0033] In step S1, the historical active power sequence p is obtained. t and satellite thermal infrared brightness temperature image sequence The target for future day prediction is the average hourly active power within the next natural day. If predicted values ​​from D+1 to D+7 are required, corresponding prediction models can be trained or called for different future day indices, or a shared principal network and different output heads can be used in an equivalent implementation.

[0034] (1) Where, p t L represents the historical hourly active power sequence prior to the prediction starting point t, including records from the current time. y This indicates the historical power playback length; in this example, 7 days, or 168 hours, are recorded. t-Ly+1 to p t This indicates the active power values ​​for each hour arranged chronologically within the playback window.

[0035] (2) Among them, y t,h H represents the actual daily average active power on the h-th natural day from the prediction starting point t, where h ranges from 1 to 7; d H represents the number of hourly records within a calendar day, as shown in the example. d =24;τ h (t) represents the index of the first hour of the h-th future natural day; u represents the hour offset within that natural day; p τh(t)+u This represents the actual active power of the target calendar day corresponding to the hour.

[0036] In step S2, the mean and standard deviation of the historical active power sequence are calculated and normalized according to formula (3). Then, the normalized sequence is divided according to segment length L. p and step size S p The system is divided into multiple overlapping hour segments. Each segment is linearly projected and its position is superimposed, forming a hydropower operation status segment token h according to formula (4). t,j This token describes local operating patterns, such as recent level changes, fluctuations, and climbs.

[0037] (3) in, p represents the normalized hourly active power. t μ represents the hourly active power before normalization. t and σ t Let represent the instance mean and instance standard deviation of the current historical input sequence, respectively. When σ t When the value is close to zero, a preset small constant can be added to the denominator to avoid numerical instability. (4) Among them, h t,j P represents the token for the j-th hydropower operation state segment corresponding to the prediction starting point t; t,j This represents the j-th power segment extracted from the normalized history sequence; j represents the segment position index; W p and b pThese represent the trainable linear projective weights and biases, respectively. This represents the learnable position embedding of the j-th segment.

[0038] In step S3, a hydropower operation state descriptor a is constructed based on the same historical active power sequence. t In this embodiment, the nearest window R is the 24 hours before the prediction start point, and the reference window B is the 24 hours immediately preceding the nearest window. The descriptor includes the nearest window mean, the nearest window standard deviation, the current active power, the difference between the first and last values ​​of the nearest window, the difference between the nearest window mean and the reference window mean, and the average absolute ramp of the nearest window. Subsequently, through... =φ q (a t ) Obtain the hydropower status query vector; φ q This represents a trainable query mapping function that sequentially includes layer normalization, linear mapping, GELU activation, and dropout.

[0039] (5) Among them, a t The six-dimensional hydropower operation status descriptor represents the prediction starting point t; R represents the most recent statistical window, in this example R = 24 hours; B represents the reference window immediately preceding R and also of length 24 hours; μ R and σ R Let p represent the mean and standard deviation of the power within window R, respectively; t p represents the active power for the current hour. t -p t-R+1 μ represents the power difference between the first and last windows in the most recent window, used to characterize short-term trends. R -μ B This indicates the mean offset of the most recent window relative to the reference window; It represents the average absolute difference in power between adjacent hours within window R, used to characterize the climbing intensity.

[0040] In step S4, valid pixel identification and invalid value filling are performed on the satellite thermal infrared brightness temperature image sequence. For the space region Ω r The cold cloud coverage is calculated according to formula (6), and the cold cloud depth is calculated according to formula (7). Effective pixel ratio, local gradient texture, global channel statistics, regional mean, cold anomaly intensity, and coverage and depth under multiple cold cloud thresholds can also be extracted. For the physical cloud event vector of each frame, first-order difference, relative mean deviation, short-window smoothing, long-window smoothing, and cumulative mean are further concatenated according to formula (8) to obtain multi-scale temporal evolution features.

[0041] (6) in, Ω represents the cold cloud coverage of the l-th frame satellite image at prediction starting point t within the r-th spatial region; r Represents the set of pixels in the r-th spatial region; u represents the pixel location; M t,l (u) is the valid pixel mask, which is 1 when valid and 0 when invalid; Indicates the thermal infrared brightness temperature of pixel u; τ c ε represents the preset cold cloud brightness temperature threshold; I(·) represents the indicator function; ε represents the preset small constant to prevent the denominator from being zero.

[0042] (7) in, This represents the depth of cold clouds in the corresponding spatial region; max(0,τ) c - ) represents the positive temperature difference when the pixel brightness temperature is lower than the cold cloud threshold, and the other symbols are the same as in formula (6). This formula normalizes and averages the positive temperature difference of effective pixels in the region, and is used to characterize the cooling intensity of the cold cloud top relative to the threshold.

[0043] (8) Where, m t,l b represents the multi-scale temporal evolution characteristics of the l-th frame satellite image; t,l This represents the original physical cloud event vector for this frame; Δb t,l This represents the first-order difference between adjacent frames; S represents the temporal mean of the frame-level cloud event vectors within the review window. s (·) and S l (·) represents the short window and smooth long window operators, respectively; C(·) represents the cumulative mean operator up to the current frame; square brackets indicate concatenation by feature dimension.

[0044] In step S5, the multi-scale TBB time evolution features are interpolated to the same length as the historical active power sequence, and the potential satellite event token u is obtained through satellite projection. t,τ Where τ is the aligned satellite time index. The result obtained in step S3... Weight W q Mapped to query vector q t And respectively through weight W k and W v will u t,τ Mapped to key vector k t,τ Sum vector v t,τ The time attention weight β is calculated according to formula (9). t,τ And obtain the satellite context guided by the hydroelectric state according to formula (10).

[0045] (9) Where, β t,τ The normalized attention weight represents the predicted satellite time position τ at the starting point t; q t Represents the query vector for hydropower status. The query vector obtained by mapping; k t,τ Indicates satellite event token u t,τ The corresponding key vector; the superscript T indicates vector transpose; τ' indicates the traversal index in the attention normalization summation; L y This represents the length of the satellite token after alignment with the historical power sequence; exp(·) represents the exponential function. Each β t,τ The sum is 1.

[0046] (10) in, This indicates the hydropower state guidance satellite context for the prediction starting point t; v t,τ Indicates satellite event token u t,τ The corresponding value vector. Then... Copy to each time location and match with u t,τ splicing, via z t,τ =φ z ([u t,τ ; ]) Obtain the hydropower state condition satellite token; where the semicolon indicates vector concatenation, φ z This represents a trainable fusion mapping function that performs normalization and linear projection on the spliced ​​result.

[0047] In step S6, via s t,τ =W s LN(z t,τ Project the hydropower state conditional satellite tokens to the same hidden dimension as the hydropower operation state fragment tokens; where LN represents layer normalization, W s This represents trainable projective weights. For each hydropower operating state segment, the token h... t,j The fragment-level cross-attention weights relative to each satellite time token are calculated according to formula (11). Then, the satellite context is aggregated according to these weights, and the aggregation result is concatenated with the corresponding hydropower operation status fragment tokens and fused by projection to form fused fragment tokens. The fused representation for prediction is then obtained by the Transformer encoder.

[0048] (11) Where, α t,j,τ h represents the cross-attention weight of the j-th hydropower segment at prediction start point t to the τ-th satellite time token; t,j This represents the token for the j-th hydropower operation status segment; s t,τIndicates the projected satellite token; W Q and W K Let τ' and L represent the trainable query projection matrix and key projection matrix in cross-attention, respectively; d represents the attention hidden dimension, and √d is used to scale the dot product; y The meaning is the same as that of formula (9).

[0049] In step S7, the prediction head outputs a normalized prediction scalar based on the fusion representation, and performs inverse normalization using the instance mean and standard deviation from step S2 to obtain the predicted daily average active power for future natural days. During the training phase, the squared error objective shown in formula (12) can be used. Training data, validation data, and test data are divided according to time boundaries. The model selection uses only internal validation samples from the training phase to avoid information leakage during testing.

[0050] (12) Where J(θ) represents the training objective function; θ represents all trainable parameters of the model; B represents a set of sample indices for a training mini-batch, and |B| represents the number of samples in that set; y t,h F represents the actual average daily active power on the h-th natural day in the future; θ (·) represents the prediction model determined by parameter θ; p t Indicates historical active power input; This represents the sequence of satellite thermal infrared brightness temperature images prior to the prediction start point. The formula trains the model by minimizing the mean squared error between the predicted and actual values.

[0051] In this embodiment, the experiment used 13,785 hourly active power records covering the period from January 29, 2024 to August 17, 2025; and 12,936 hourly FY-4B satellite TBB images covering the period from March 5, 2024 to August 26, 2025, with a quality-controlled image frame availability of 97.7%. The main satellite input consisted of a 75x75 grid of mesoscale contextual image patches composed of four TBB-derived channels. The fixed monthly reserved test set included the high-water season in September 2024, the normal-water season in November 2024, and the low-water season in February 2025. The results in the table are the average of three random seeds.

[0052] The RMSE of the complete model on D+1 to D+7 were 2.160, 2.436, 2.422, 2.437, 2.481, 2.495 and 2.648 respectively, with corresponding MAPE of 19.641%, 24.379%, 29.038%, 31.121%, 34.365%, 36.519% and 38.754% respectively; the average RMSE for the seven forecast days was 2.440 and the average MAPE was 30.545%. Under the same data partitioning and input window, the complete model achieved the lowest RMSE from D+2 to D+7, and the lowest MAPE at D+3, D+4, and D+7. Its average RMSE decreased from 2.560 to 2.440 compared to the historical PatchTST baseline with the best average RMSE, a reduction of approximately 4.69%. Its average MAPE decreased from 31.369% to 30.545% compared to the TimeMixer baseline with the best average MAPE, a reduction of approximately 2.63%. The above data are used to illustrate the technical effects of this invention in predicting hydropower active power over multiple forecast days, and do not limit the scope of protection of the claims.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the power output of hydropower stations based on satellite thermal infrared cloud event fusion, characterized in that, Includes the following steps: S1. Obtain the historical active power sequence before the prediction start point and the satellite thermal infrared brightness temperature image sequence within the corresponding time range, and determine the future natural days to be predicted; S2. The historical active power sequence is normalized and divided into overlapping hour segments, and each overlapping hour segment is mapped to a hydropower operation status segment token containing a location code. S3. Construct a hydropower operation status descriptor based on the nearest window of the historical active power sequence, and map the hydropower operation status descriptor into a hydropower status query vector; S4. Perform effective pixel processing on the satellite thermal infrared brightness temperature image sequence, and extract cold cloud coverage, cold cloud depth, spatial texture, regional statistics and temporal evolution features from the brightness temperature image to form a time-aligned satellite cloud event token sequence. S5. Perform time attention on the satellite cloud event token sequence using the hydropower status query vector to obtain the satellite context guided by hydropower status, and merge the satellite context guided by hydropower status with the satellite cloud event tokens at each time moment into a hydropower status conditional satellite token. S6. Using each hydropower operation status segment token as a query, perform segment-level cross-attention on the hydropower status conditional satellite token to obtain the satellite context corresponding to each hydropower operation status segment token, merge the corresponding hydropower operation status segment token and satellite context into a fused segment token, and encode the fused segment token. S7. Output the predicted daily average active power for the future natural days based on the encoding results.

2. The hydropower station power prediction method based on satellite thermal infrared cloud event fusion as described in claim 1, characterized in that, The historical active power sequence includes active power records of the power plant for several consecutive hours prior to the prediction start point, the satellite thermal infrared brightness temperature image sequence includes multi-channel thermal infrared brightness temperature image blocks prior to the prediction start point, and the daily average active power prediction value is the average of the active power records for several hours within the future natural day.

3. The hydropower station power prediction method based on satellite thermal infrared cloud event fusion as described in claim 1, characterized in that, In step S2, instance normalization is performed based on the mean and variance of the historical active power sequence; Multiple overlapping hourly segments are formed according to a preset segment length and a preset step size; each overlapping hourly segment is linearly mapped and superimposed with a learnable position embedding to obtain the hydropower operation status segment token.

4. The hydropower station power prediction method based on satellite thermal infrared cloud event fusion as described in claim 1, characterized in that, The hydropower operation status descriptor includes the mean of the most recent window, the standard deviation of the most recent window, the active power at the prediction starting point, the difference between the active power at the beginning and end of the most recent window, the difference between the mean of the most recent window and the mean of the previous reference window, and the average absolute ramp of the most recent window; wherein, the most recent window is the sample of the 24 hours before the prediction starting point, and the previous reference window is the sample of the 24 hours immediately preceding the most recent window.

5. The hydropower station power prediction method based on satellite thermal infrared cloud event fusion as described in claim 1, characterized in that, In step S4, the effective pixels of each satellite thermal infrared brightness temperature image are identified; within a preset spatial partition, the cold cloud coverage is determined based on the proportion of effective pixels below the cold cloud brightness temperature threshold, and the cold cloud depth is determined based on the positive difference between the cold cloud brightness temperature threshold and the effective pixel brightness temperature; at least one cloud event descriptor is also extracted from the effective pixel proportion, local spatial texture, global channel statistics, regional mean, and cold anomaly intensity.

6. The hydropower station power prediction method based on satellite thermal infrared cloud event fusion as described in claim 5, characterized in that, The time evolution features include the first-order difference of the satellite cloud event descriptor, the deviation relative to the retrospective mean, the short-window smoothing result, the long-window smoothing result, and the cumulative mean result, and the time evolution features are interpolated to the same time length as the historical active power sequence.

7. The hydropower station power prediction method based on satellite thermal infrared cloud event fusion as described in claim 1, characterized in that, In step S5, the hydropower status query vector is mapped to a query vector, and the satellite cloud event token is mapped to a key vector and a value vector; the time attention weight is calculated based on the similarity between the query vector and each key vector. The satellite context guided by the hydroelectric state is obtained by weighting and summing the value vectors according to the time attention weights.

8. The hydropower station power prediction method based on satellite thermal infrared cloud event fusion as described in claim 1, characterized in that, In step S6, the hydropower status conditional satellite token is projected to the same hidden dimension as the hydropower operation status fragment token; For each hydropower operation status segment token, calculate its cross-attention weight with each hydropower status conditional satellite token; The satellite context corresponding to the hydropower operation status fragment token is obtained according to the cross-attention weight, and the satellite context and the hydropower operation status fragment token are concatenated and then fused and projected.

9. The hydropower station power prediction method based on satellite thermal infrared cloud event fusion as described in any one of claims 1 to 8, characterized in that, After encoding the fused fragment token, a normalized prediction scalar is obtained through the prediction head, and the normalized prediction scalar is denormalized using the mean and standard deviation of the historical active power sequence to obtain the predicted value of the daily average active power for the future natural days. During training, the squared error between the predicted value and the actual future daily average active power is used as the training loss.

10. A hydropower station power prediction system based on satellite thermal infrared cloud event fusion, characterized in that, include: Data acquisition module: acquires the historical active power sequence before the prediction start point and the satellite thermal infrared brightness temperature image sequence within the corresponding time range, and determines the future natural days to be predicted; Hydropower Segment Token Module: Normalizes the historical active power sequence into instances and divides it into overlapping hour segments, mapping each overlapping hour segment into a hydropower operation status segment token containing a location code; Hydropower status query module: Constructs a hydropower operation status descriptor based on the most recent window of the historical active power sequence, and maps the hydropower operation status descriptor to a hydropower status query vector; Satellite cloud event token module: Performs effective pixel processing on the satellite thermal infrared brightness temperature image sequence, and extracts cold cloud coverage, cold cloud depth, spatial texture, regional statistics and temporal evolution features from the brightness temperature image to form a time-aligned satellite cloud event token sequence; Hydropower guidance context module: Utilizes the hydropower status query vector to perform time attention on the satellite cloud event token sequence to obtain the hydropower status-guided satellite context, and merges the hydropower status-guided satellite context with the satellite cloud event tokens at each time moment into a hydropower status-conditionalized satellite token; Fragment-level cross-fusion module: Using each hydropower operation status fragment token as a query, perform fragment-level cross-attention on the hydropower status conditional satellite token to obtain the satellite context corresponding to each hydropower operation status fragment token, fuse the corresponding hydropower operation status fragment token and satellite context into a fused fragment token, and encode the fused fragment token; Power prediction module: Outputs the predicted daily average active power for the next natural day based on the encoding results.

Images