Method and system for predicting power generation capacity of hydropower station

Abstract

The invention discloses a hydropower station power generation capacity prediction method and system, and relates to the hydropower dispatching and data processing technology, and the method comprises the steps: combining weather text data and power generation data to obtain a correlation description text; intercepting a plurality of pieces of image data corresponding to the divided time periods; power generation text features are extracted, and water level image features are extracted; utilizing a long short-term memory (LSTM) network to obtain a power generation text feature of the context information and a water level feature of the global information; performing weighted sum on the power generation text features of the context information and the water level features of the global information by using a cross-modal bilinear attention fusion mechanism; inputting the power generation characteristics associated with the water level and the power generation characteristics extracted based on the power generation data into a Transformer decoder for decoding, and converting the power generation characteristics into water level probability distribution behind the dam of the hydropower station by using Softmax; and executing generating capacity prediction according to the water level probability distribution behind the dam of the hydropower station. The power generation capacity of the hydropower station is predicted in combination with the image-text mode, and the accuracy of power generation prediction is improved.

Description

Technical Field

[0001] The present application relates to the technical field of hydropower dispatching and data processing, and in particular to a method and system for predicting the power generation capacity of a hydropower station. Background Art

[0002] Hydropower forecasting plays a vital role in energy management and planning. As the proportion of renewable energy in the energy mix continues to increase, the reliability and stability of hydropower supply, as a clean and renewable energy form, is crucial to the operation of the power system. The accuracy of hydropower forecasting directly affects the operational safety, economy and sustainability of the power system.

[0003] Hydropower forecasting methods can be roughly divided into two categories: physical model-based and data-driven methods. The physical model-based method predicts hydropower generation by establishing a hydrological and hydraulic model, taking into account factors such as reservoir water level, river flow, and rainfall. By sampling hydropower measurement data, physical empirical formulas are used to calculate hydropower output. The data-driven method directly uses historical data for modeling and forecasting. For example, the existing technology uses EMD (empirical mode decomposition) to decompose the inflow flow sequence, and then establishes LSTM (long short term memory) models for each decomposed component for forecasting. Existing methods need to fully consider the complexity of hydrological and hydraulic processes, but may perform poorly when data is insufficient or hydrological conditions change drastically. Summary of the invention

[0004] The embodiment of the present application combines the associated description text related to power generation and the post-dam monitoring video data, and through cross-modal feature fusion, realizes the prediction of the power generation capacity of the hydropower station by combining pictures and texts, thereby improving the accuracy of power generation prediction.

[0005] The present application embodiment provides a method for predicting the power generation capacity of a hydropower station, including: Acquire weather text data from a target weather website and obtain power generation data of power generators by time period, combine the weather text data and power generation data to obtain associated description text; and, Obtain video surveillance data behind the hydropower station dam and capture multiple image data corresponding to the time periods; Extracting power generation text features from the associated description text, and extracting water level image features from the plurality of image data; The extracted power generation text features and water level image features are respectively used in a long short-term memory network (LSTM) to obtain the power generation text features of contextual information and the water level features of global information; The power generation text features of contextual information and the water level features of global information are weighted and summed using a cross-modal bilinear attention fusion mechanism; The weighted results are passed through the Transformer encoder to obtain the power generation characteristics associated with the water level; The power generation features associated with the water level and the power generation features extracted based on the power generation data are input into the Transformer decoder for decoding, and the decoding results are mapped into a vector of the water level relationship dimension; According to the vector of the mapped water level relationship dimension, Softmax is used to convert it into the probability distribution of the water level behind the hydropower station dam; Power generation forecasting is performed based on the probability distribution of water levels behind the dam of a hydropower station.

[0006] Optionally, obtaining weather text data from a target weather website and obtaining power generation data of a generator set by time period include: Determine the update interval of the target weather website and the load dispatch points of the hydropower station generator units; Combining weather text data and power generation data to obtain associated description text includes: Mapping the load dispatch points of the generator set to the corresponding time periods of the update time interval of the target weather website; The weather text data and power generation data of each time period are combined to obtain the associated description text of the corresponding time period.

[0007] Optionally, obtaining video surveillance data behind the hydropower station dam and intercepting multiple image data corresponding to the time periods include: The video surveillance data behind the hydropower station dam is divided into multiple sub-video segments according to each time period, and estimating ambient lighting conditions of a plurality of sub-video segments according to time intervals of each divided time period; For a sub-video segment with good ambient lighting conditions, multiple frames of global images are extracted from the sub-video segment according to a set step size; For sub-video segments with poor ambient lighting conditions, multiple frames of global images are randomly extracted from the sub-video segments and local images under a fixed lighting range are selected from the multiple frames of global images using a preset image selection area.

[0008] Optionally, extracting power generation text features from the associated description text includes: Segmenting the associated description text, and matching each segmented word with a plurality of preset parameter types; Obtain description and numerical data within the segmentation distance before and after the matching position from the associated description text; The features of the power generation text are extracted based on the description and numerical data within the segmentation distance before and after the matching position.

[0009] Optionally, extracting water level image features from multiple image data includes: Predetermine multiple water level lines within the dam area based on dam area reference objects; Identify the dam area and water area in each global image or local image; For the dam area and water area in the identified global image and local image, matching is performed based on the dam area reference object to map the water area boundary to a water level line or between two water level lines; Based on the mapped water level lines and the identified dam area and water area in the global image or local image, water level image features are extracted.

[0010] Optionally, the extracted power generation text features and water level image features are respectively used to obtain the power generation text features of contextual information and the water level features of global information to satisfy: in, , are the hidden states corresponding to the text features of the i-th step and the i-1-th step, respectively. , are the cell states corresponding to the text features of the i-th step and the i-1-th step, respectively. is the power generation text feature of the i-th step; in, , are the hidden states corresponding to the water level image features of the i-th step and the i-1-th step, respectively. , are the cell states corresponding to the water level image features of the i-th step and the i-1-th step, respectively. is the water level image feature of the i-th step.

[0011] Optionally, the power generation text features of the context information and the water level features of the global information are weighted and combined using a cross-modal bilinear attention fusion mechanism including: in, , is the weighting parameter, is the output of LSTM based on water level image features, is the output of LSTM based on the features of the generated text, CMA() is the cross-modal multi-head attention, is the softmax function, is the weight matrix of water level image features, , are the first and second weight matrices of the power generation text features; Z is the bilinear weight matrix, , is the trainable weight + =1, is the weighted sum.

[0012] Optionally, the cross entropy loss between the probability distribution predicted by the model and the distribution based on historical water level data is calculated during the training process.

[0013] Optionally, performing power generation forecasting based on the probability distribution of the water level behind the hydropower station dam includes: According to the continuously predicted probability distribution of water level behind the dam of the hydropower station, the dispatchable power generation limit of the corresponding hydropower station unit is determined to realize the prediction of power generation capacity.

[0014] The embodiment of the present application also proposes a hydropower station power generation capacity prediction system, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the hydropower station power generation capacity prediction method as described above are implemented.

[0015] The method of the embodiment of the present application combines the associated description text related to power generation and the post-dam monitoring video data, and through cross-modal feature fusion, realizes the prediction of the power generation capacity of the hydropower station in a combined manner of text and graphics, thereby improving the accuracy of power generation prediction.

[0016] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 The basic process diagram of the method for predicting the power generation capacity of a hydropower station according to the embodiment of the present application is shown in FIG. DETAILED DESCRIPTION

[0018] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0019] The present application embodiment proposes a method for predicting the power generation capacity of a hydropower station, such as Figure 1 As shown, the following steps are included: In step S101, weather text data and power generation data of generator sets are obtained from a target weather website in different time periods, and the weather text data and power generation data are combined to obtain associated description text. Specifically, a crawler tool can be used to crawl the weather website according to the update time interval of the weather website and set a corresponding time step. Weather text data at multiple time points can be crawled in one day, and then the actual power generation data of generator sets under multiple scheduling nodes can be obtained, so as to form an associated description text.

[0020] In step S102, video surveillance data behind the hydropower station dam is obtained, and multiple image data corresponding to the time periods are intercepted. The present application further and more specifically combines the actual water level changes in the video surveillance data behind the dam to achieve power generation prediction from multiple features.

[0021] In step S103, power generation text features are extracted from the associated description text, and water level image features are extracted from the plurality of image data.

[0022] In step S104, the extracted power generation text features and water level image features are respectively used to obtain power generation text features of contextual information and water level features of global information using a long short-term memory network LSTM.

[0023] In step S105, the power generation text features of the context information and the water level features of the global information are weighted and added using a cross-modal bilinear attention fusion mechanism. In a specific example, the embodiment of the present application performs feature fusion through bilinear attention of the text modality and the image modality, which can integrate the water level features represented by the weather, power generation, and monitoring images, and further complete the prediction based on the fused features in the subsequent steps.

[0024] In step S106, the weighted result is passed through a Transformer encoder to obtain power generation characteristics associated with the water level.

[0025] In step S107, the power generation characteristics associated with the water level and the power generation characteristics extracted based on the power generation data are input into the Transformer decoder for decoding, and the decoding result is mapped into a vector of the water level relationship dimension.

[0026] In step S108, the vector of the mapped water level relationship dimension is converted into a water level probability distribution behind the hydropower station dam using Softmax.

[0027] In step S109, power generation prediction is performed based on the water level probability distribution behind the dam of the hydropower station.

[0028] In a specific embodiment, obtaining weather text data from a target weather website and obtaining power generation data of a generator set by time period includes: Determine the update time interval of the target weather website and the load dispatching points of the hydropower station generator set. In some embodiments, for example, some weather websites are updated every 15 minutes, 20 minutes, and 1 hour, and the load dispatching points of the hydropower station generator set are, for example, 96 points, that is, every 15 minutes.

[0029] Combining weather text data and power generation data to obtain associated description text includes: The load dispatching points of the generator set are mapped to the corresponding time periods of the update time interval of the target weather website. Specifically, the 96 points can be mapped to the update time intervals of the weather website according to the update interval of the weather website.

[0030] The weather text data and power generation data of each time period are combined to obtain the associated description text of the corresponding time period. That is, the data is combined according to the aforementioned weather update time period to obtain the associated description text. In the specific example, the power generation data can include the power generation log, so that the associated description text contains the weather data and the power generation data of the generator set.

[0031] In some embodiments, obtaining video surveillance data behind a hydropower station dam and capturing multiple image data corresponding to time periods includes: The video surveillance data behind the hydropower station dam is divided into multiple sub-video segments according to each time period, and The ambient light conditions of the plurality of sub-video segments are estimated according to the time interval of each time period. In a specific example, for example, the ambient light is divided into good light conditions from 08:00 to 18:00 every day and dark light conditions at other times according to the season.

[0032] For sub-video segments with good ambient lighting conditions, multiple frames of global images are extracted from the sub-video segments according to the set step size. For sub-video segments with poor ambient lighting conditions, multiple frames of global images are randomly extracted from the sub-video segments and a local image under a fixed lighting range is selected from the multiple frames of global images using a preset image selection area. In a specific example, for poor lighting conditions at night, a local image with lighting conditions behind the dam is selected to facilitate the determination of water level features in the image. The method of the embodiment of the present application utilizes the water level features in the image to solve the problem of inaccurate power generation prediction when hydrological conditions change dramatically, and the combination of weather text and power generation features can further improve the accuracy of the prediction.

[0033] In some embodiments, extracting power generation text features from the associated description text includes: The associated description text is segmented, and each segmentation is matched with a plurality of preset parameter types. In some embodiments, for example, the weather text may contain air quality 61, wind speed northwest wind, humidity 51%, visibility 6.6 kilometers, air pressure 991 hPa, and light rain is expected tomorrow. So far, today's precipitation is not much different from yesterday. The crawled weather text is matched by the set parameter types such as rainfall, wind speed, etc., and the parameter unit can also be used to match, so as to match the weather text associated with power generation.

[0034] The description and numerical data within the segmentation distance before and after the matching position are obtained from the associated description text. For example, for precipitation, according to the above example, "Today's precipitation is not much different from yesterday's" is crawled, and the matching process is similar for other texts.

[0035] The features of the power generation text are extracted based on the description and numerical data within the segmentation distance before and after the matching position.

[0036] In some embodiments, extracting water level image features from multiple image data includes: A plurality of water level lines are determined in advance within the dam area based on reference objects in the dam area. For example, a plurality of horizontal water level lines with intervals of 0.3 m or 0.5 m can be determined for the dam body based on actual reference objects on the dam body. The reference objects can be fixed objects on the dam body. The specific intervals can be determined based on the resolution of the actual monitoring equipment. In subsequent examples, the water level lines are mapped to global images or local images.

[0037] Identify the dam area and water area in each global image or local image.

[0038] For the dam area and water area in the identified global image and local image, matching is performed based on the dam area reference object to map the water area boundary to a water level line or between two water level lines. In some examples, if mapped between two water level lines, the actual water level range can be determined based on the pixel ratio relationship. That is, through image features, the problem of distortion of hydrological monitoring data when hydrological conditions change drastically can be solved.

[0039] Based on the mapped water level lines and the identified dam area and water area in the global image or local image, water level image features are extracted.

[0040] In some embodiments, the extracted power generation text features and water level image features are respectively used to obtain the power generation text features of context information and the water level features of global information using a long short-term memory network LSTM to satisfy: in, , are the hidden states corresponding to the text features of the i-th step and the i-1-th step, respectively. , are the cell states corresponding to the text features of the i-th step and the i-1-th step, respectively. is the power generation text feature of the i-th step; in, , are the hidden states corresponding to the water level image features of the i-th step and the i-1-th step, respectively. , are the cell states corresponding to the water level image features of the i-th step and the i-1-th step, respectively. is the water level image feature of the i-th step. In the specific example, the input to LSTM is the extracted power generation text feature sequence and the water level image feature sequence, that is, the combination of multiple features of continuous time periods, so as to obtain the power generation text feature of contextual information and the water level feature of global information.

[0041] In some embodiments, the power generation text features of the context information and the water level features of the global information are weighted and combined using a cross-modal bilinear attention fusion mechanism, including: in, , is the weighting parameter, is the output of LSTM based on water level image features, is the output of LSTM based on the features of the generated text, CMA() is the cross-modal multi-head attention, is the softmax function, is the weight matrix of water level image features, , are the first and second weight matrices of the power generation text features; Z is the bilinear weight matrix, , is the trainable weight + =1, is the weighted sum.

[0042] In some embodiments, the method further includes calculating the cross entropy loss between the probability distribution predicted by the model and the distribution based on historical water level data during the training process.

[0043] In some embodiments, performing power generation prediction based on the water level probability distribution behind the hydropower station dam includes: determining the dispatchable power generation limit of the corresponding hydropower station unit based on the continuously predicted water level probability distribution behind the hydropower station dam to achieve power generation capacity prediction. In a specific example, based on the continuously predicted water level probability distribution behind the hydropower station dam, the water level with the largest probability can be selected to perform continuous available power generation capacity prediction.

[0044] The method of the embodiment of the present application combines the associated description text related to power generation and the post-dam monitoring video data, and through cross-modal feature fusion, realizes the prediction of the power generation capacity of the hydropower station in a combined manner of text and graphics, thereby improving the accuracy of power generation prediction.

[0045] The embodiment of the present application also proposes a hydropower station power generation capacity prediction system, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the hydropower station power generation capacity prediction method as described above are implemented.

[0046] In addition, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present disclosure with equivalent elements, modifications, omissions, combinations (e.g., various embodiments intersecting schemes), adaptations or changes. It is not limited to the examples described in this specification or during the implementation of this application, and its examples will be interpreted as non-exclusive.

[0047] The above description is intended to be illustrative rather than restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. For example, those of ordinary skill in the art may use other embodiments when reading the above description.

[0048] The above embodiments are merely exemplary embodiments of the present disclosure. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the essence and protection scope of the present disclosure, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present invention.

Claims

1. A method for predicting the power generation capacity of a hydropower station, characterized in that: include: Acquire weather text data from a target weather website and obtain power generation data of power generators by time period, and combine the weather text data and power generation data to obtain associated description text; as well as, Obtain video surveillance data behind the hydropower station dam and capture multiple image data corresponding to the time periods; Extracting power generation text features from the associated description text, and extracting water level image features from the plurality of image data; The extracted power generation text features and water level image features are respectively used in a long short-term memory network (LSTM) to obtain the power generation text features of contextual information and the water level features of global information; The power generation text features of contextual information and the water level features of global information are weighted and summed using a cross-modal bilinear attention fusion mechanism; The weighted results are passed through the Transformer encoder to obtain the power generation characteristics associated with the water level; The power generation features associated with the water level and the power generation features extracted based on the power generation data are input into the Transformer decoder for decoding, and the decoding results are mapped into a vector of the water level relationship dimension; According to the vector of the mapped water level relationship dimension, Softmax is used to convert it into the probability distribution of the water level behind the hydropower station dam; Power generation forecasting is performed based on the probability distribution of water levels behind the dam of a hydropower station.

2. The method for predicting the power generation capacity of a hydropower station according to claim 1, characterized in that: Obtaining weather text data from the target weather website and obtaining power generation data of generator sets by time period includes: Determine the update interval of the target weather website and the load dispatch points of the hydropower station generator units; Combining weather text data and power generation data to obtain associated description text includes: Mapping the load dispatch points of the generator set to the corresponding time periods of the update time interval of the target weather website; The weather text data and power generation data of each time period are combined to obtain the associated description text of the corresponding time period.

3. The method for predicting the power generation capacity of a hydropower station according to claim 2, characterized in that: Obtain video surveillance data behind the hydropower station dam, and intercept multiple image data corresponding to the time periods, including: The video surveillance data behind the hydropower station dam is divided into multiple sub-video segments according to each time period, and estimating ambient lighting conditions of a plurality of sub-video segments according to time intervals of each divided time period; For a sub-video segment with good ambient lighting conditions, multiple frames of global images are extracted from the sub-video segment according to a set step size; For sub-video segments with poor ambient lighting conditions, multiple frames of global images are randomly extracted from the sub-video segments and local images under a fixed lighting range are selected from the multiple frames of global images using a preset image selection area.

4. The method for predicting the power generation capacity of a hydropower station according to claim 3, characterized in that: Extracting power generation text features from the associated description text includes: Segmenting the associated description text, and matching each segmented word with a plurality of preset parameter types; Obtain description and numerical data within the segmentation distance before and after the matching position from the associated description text; The features of the power generation text are extracted based on the description and numerical data within the segmentation distance before and after the matching position.

5. The method for predicting the power generation capacity of a hydropower station according to claim 3, characterized in that: Extracting water level image features from multiple image data includes: Predetermine multiple water level lines within the dam area based on dam area reference objects; Identify the dam area and water area in each global image or local image; For the dam area and water area in the identified global image and local image, matching is performed based on the dam area reference object to map the water area boundary to a water level line or between two water level lines; Based on the mapped water level lines and the identified dam area and water area in the global image or local image, water level image features are extracted.

6. The method for predicting the power generation capacity of a hydropower station according to claim 3, characterized in that: The extracted power generation text features and water level image features are respectively used in the long short-term memory network LSTM to obtain the power generation text features of contextual information and the water level features of global information to meet the following requirements: in, , are the hidden states corresponding to the text features of the i-th step and the i-1-th step, respectively. , are the cell states corresponding to the text features of the i-th step and the i-1-th step, respectively. is the power generation text feature of the i-th step; in, , are the hidden states corresponding to the water level image features of the i-th step and the i-1-th step, respectively. , are the cell states corresponding to the water level image features of the i-th step and the i-1-th step, respectively. is the water level image feature of the i-th step.

7. The method for predicting the power generation capacity of a hydropower station according to claim 6, characterized in that: The power generation text features of contextual information and the water level features of global information are weighted and fused using the cross-modal bilinear attention mechanism, including: in, , is the weighting parameter, is the output of LSTM based on water level image features, is the output of LSTM based on the features of the generated text, CMA() is the cross-modal multi-head attention, is the softmax function, is the weight matrix of water level image features, , are the first and second weight matrices of the power generation text features; Z is the bilinear weight matrix, , is the trainable weight + =1, is the weighted sum.

8. The method for predicting the power generation capacity of a hydropower station according to claim 7, characterized in that: It also includes calculating the cross entropy loss between the probability distribution predicted by the model and the distribution based on historical water level data during the training process.

9. The method for predicting the power generation capacity of a hydropower station according to claim 7, characterized in that: Power generation forecasting based on the probability distribution of water levels behind the hydropower station dam includes: According to the continuously predicted probability distribution of water level behind the dam of the hydropower station, the dispatchable power generation limit of the corresponding hydropower station unit is determined to realize the prediction of power generation capacity.

10. A hydropower station power generation capacity prediction system, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method for predicting the power generation capacity of a hydropower station as claimed in any one of claims 1 to 9 are implemented.

Images