Radial flow type hydropower station generating capacity prediction method based on depth data fusion

Abstract

The invention relates to a runoff hydropower station generating capacity prediction method based on depth data fusion, and the method comprises the following steps: S1, carrying out data layer fusion, integrating multi-source hydro meteorological data, obtaining multi-source data from a meteorological station, a hydrological station, a satellite remote sensing platform and a numerical weather forecasting system, carrying out the cleaning, interpolation and normalization processing of the data, and carrying out the prediction of the generating capacity of a runoff hydropower station; performing data space-time alignment; s2, carrying out feature layer fusion, firstly inputting a multi-source data set of a unified space-time reference, carrying out feature extraction, carrying out feature fusion based on a self-attention mechanism, embedding a water balance equation in a deep learning model, and obtaining a comprehensive feature vector; and S3, performing prediction, inputting the comprehensive feature vector in the step S2, and performing power generation prediction by using a deep neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower energy prediction, and particularly to a method for predicting the power generation of a run-of-river hydropower station based on deep data fusion, which is applicable to the power generation prediction of run-of-river hydropower stations. Background Art

[0002] A run-of-river hydropower station is a facility that uses the natural flow of a river for power generation. It usually does not have a large reservoir or only has a small water storage capacity, and is an important part of clean energy. However, due to the lack of large-scale water storage capacity, the power generation of a run-of-river hydropower station highly depends on the natural flow of the river, which makes the prediction of power generation particularly important. Accurate power generation prediction not only helps the dispatching and planning of the power grid, but also improves the economic efficiency and reliability of power production.

[0003] Firstly, the power generation of a run-of-river hydropower station highly depends on natural water inflow conditions, and its uncertainty is the main challenge for prediction. Different from conventional reservoir power stations, run-of-river power stations lack regulation ability, and their power generation is directly affected by the real-time changes of upstream water inflow, and is significantly affected by meteorological factors such as precipitation, temperature, evaporation, and the hydrological characteristics of the basin, resulting in large fluctuations in power generation and significantly increasing the prediction difficulty. Traditional prediction methods mainly use statistical models, such as time series analysis, regression analysis, and models based on the basin hydrological process. These methods are easy to operate, but often rely on historical data and are difficult to capture complex non-linear relationships under climate change, thus limiting the prediction accuracy.

[0004] Secondly, the power generation of a run-of-river hydropower station is affected by multiple factors such as meteorological conditions, basin runoff, and reservoir operation strategies, and has strong non-linearity and time-variability. Traditional prediction methods are difficult to fully explore the complex non-linear relationships among meteorological, hydrological, and unit operation data, resulting in limited prediction accuracy. Traditional prediction methods, such as ARIMA and SVM, rely on artificial feature engineering and are difficult to capture the dynamic associations among multi-source data. In recent years, deep learning technologies, such as LSTM and TCN, have shown advantages in time series prediction, but data scarcity and multi-modal fusion are still bottlenecks. Therefore, it is necessary to effectively integrate these multi-source data and construct a high-precision power generation prediction model.

[0005] Finally, although the existing technology combines multiple linear regression with least squares support vector machine to improve the accuracy of annual power generation prediction by using the historical data of small run-of-river hydropower stations. There are also those who use a long short-term memory network (LSTM) model to construct training samples after performing stationarity and correlation tests on time series, achieving a relatively high prediction effect. However, there are still problems such as poor data quality, poor model generalization ability, difficulty in fully exploring the complex non-linear relationships among meteorological, hydrological, and unit operation data, and insufficient prediction ability. Summary of the Invention

[0006] To address the deficiencies in the existing technology, the objective of the present invention is to solve the above-mentioned defects, and further propose a runoff hydropower station power generation prediction method based on deep data fusion. Through multi-level data fusion technology, multi-source hydrometeorological data is integrated to construct a high-precision power generation prediction model, solving the problems of low prediction accuracy and poor adaptability of traditional methods.

[0007] The present invention adopts the following technical solutions.

[0008] A runoff hydropower station power generation prediction method based on deep data fusion includes the following steps:

[0009] Step S1: Perform data layer fusion, integrate multi-source hydrometeorological data, obtain multi-source data from weather stations, hydrological stations, satellite remote sensing platforms, and numerical weather prediction systems, and perform data cleaning, interpolation, normalization, and data spatio-temporal alignment.

[0010] Step S2: Perform feature layer fusion. First, input a multi-source data set with a unified spatio-temporal reference, perform feature extraction, perform feature fusion based on the self-attention mechanism, embed the water balance equation in the deep learning model, and obtain a comprehensive feature vector.

[0011] Step S3: Perform prediction. Input the comprehensive feature vector in Step S2 and use a deep neural network to predict the power generation.

[0012] Further, in Step S1, it specifically includes:

[0013] Step S11: Data spatio-temporal alignment:

[0014] Step S111: Align the satellite remote sensing data and the ground monitoring data spatially through the inverse distance weighted method:

[0015]

[0016] Among them, Z(s0) is the target grid value, d(s0, s i ) is the distance from the station to the grid center, and p is the power parameter.

[0017] Step S112: Perform time alignment through the sliding window aggregation method, aggregate the satellite remote sensing data to the same granularity as the daily ground monitoring data, input the original time series data {(t i , v i )} and the window parameter window size T. For each target time point t' j , define the time window range as [t' j - T, t' j , extract all the original data points {v i} within the window, satisfying t'j -T ≤ t i ≤ t' j , calculate the aggregated value Output the aligned time series

[0018] Step S12, train a GAN model composed of a generator and a discriminator, input the obtained historical hydro-meteorological data, generate synthetic data by the generator, and output a multi-source dataset with a unified spatio-temporal benchmark:

[0019]

[0020] Among them, D(x) is the output of the discriminator for real data, and D(G(z)) is the output of the discriminator for generated data.

[0021] Furthermore, the specific steps of step S2 include:

[0022] Step S21, input the multi-source dataset with a unified spatio-temporal benchmark, extract the spatial features of satellite remote sensing data through 3D-CNN, extract the temporal features of ground time series data through LSTM, and extract the meteorological trend features of numerical weather prediction data through Transformer, specifically including:

[0023] Step S211, input satellite remote sensing data Belonging to spatial features, where H×W is the spatial resolution, D is the number of channels, T is the time step, and the spatial feature tensor is extracted through multi-layer convolution and pooling Where T' is the time step, h×w is the spatial dimension, and d is the number of feature channels.

[0024] Step S212, input ground time series data Belonging to temporal features, where T is the time step, F is the feature dimension, and the temporal dependence relationships such as flow rate and temperature are captured through LSTM units to obtain the temporal feature vector

[0025] Step S213, input numerical weather prediction data Belonging to meteorological trend features, where T is the time step, C is the number of meteorological variables, and the long-range dependence relationships of meteorological variables are captured through the multi-head attention mechanism to obtain the meteorological trend feature vector

[0026] Step S22, dynamically fuse features based on the attention mechanism, specifically including:

[0027] Step S221, calculate the attention scores, and use the self-attention mechanism to calculate the mutual influence between features; first project each modal feature H i onto the query, key, and value spaces:

[0028] Q i = W Q H i , K i = W K H i , V i = W V H i

[0029] Among them, W Q , W K , W V is a learnable projection matrix.

[0030] Calculate the attention score where d is a scaling factor for the feature dimension.

[0031] Normalize the attention scores

[0032] Step S222, calculate the fused feature H fused , using the attention weight α ij to perform a weighted sum on the value vector V j :

[0033] H fused = ∑ j α ij V j .

[0034] Step S23, embed the water balance equation, including:

[0035] Step S231, input the fused feature vector obtained in step S222 and perform an initial flow prediction through a fully connected network

[0036] Step S232, define the water balance equation:

[0037]

[0038] Among them, is the inflow, ΔS (t) is the change in water storage, and the physical correction term ∈ is dynamically learned through the attention mechanism (t) , and the output flow prediction value is:

[0039]

[0040] The water balance equation constraint ensures that the predicted value conforms to the physical law, avoids the risk of overfitting driven by pure data, and the flow prediction result can be used as the ground monitoring data input in step S1, improving the accuracy of hydropower station power generation prediction.

[0041] Further, in step S3, the DNN model architecture and training specifically include:

[0042] Step S31, establish a DNN model framework, perform forward propagation, and calculate the hidden layer H i and the output layer

[0043] The input layer inputs the fused feature vector H obtained in step S222 fused , with dimension d. The hidden layer includes multiple fully connected layers, the activation function ReLU, and a batch normalization Dropout layer. The i-th hidden layer:

[0044] H i = σ(W i H i-1 + b i )

[0045] where H i-1 is the output of the previous layer, the first layer inputs H fused , W i , b i are trainable parameters, and σ(·) is the activation function ReLU.

[0046] The output layer performs a linear transformation to output the predicted value:

[0047]

[0048] where W out , b out are the output layer parameters, and H n is the output of the last hidden layer.

[0049] Step S32, train the DNN model, and calculate the loss function loss:

[0050]

[0051] where Y i is the actual power generation, is the power generation predicted by the DNN model.

[0052] The Adam function is used as the optimizer for training, and the optimization objective is to update W and b to minimize the loss loss.

[0053] The beneficial effects of the present invention are that, compared with the prior art, the present invention has the following advantages:

[0054] The present invention proposes a method for predicting the power generation of a run-of-river hydropower station based on deep data fusion. The inverse distance weighted interpolation IDW method is used to spatially align satellite remote sensing data with ground monitoring data, and the sliding window aggregation method is used to perform temporal alignment to ensure a unified spatiotemporal benchmark for multi-source data. Generative adversarial networks (GANs) are used to generate synthetic data, which can effectively alleviate the problem of insufficient data. At the same time, its adversarial training mechanism can enhance the modeling ability of the model for complex distributions. Based on the dynamic fusion features of the attention mechanism, the complex interactions of hydrological, meteorological, remote sensing and other data are mined to improve prediction accuracy and reduce computational complexity. The water balance equation is embedded, and the data-driven and physics-based coupling models improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the power generation prediction method of run-of-river hydropower station based on deep data fusion;

[0056] Figure 2 This is a flowchart of the power generation prediction method for run-of-river hydropower stations based on deep data fusion. DETAILED DESCRIPTION

[0057] The present application is further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present application.

[0058] As an efficient, clean and renewable energy source, hydropower occupies an important position in the global energy structure. In the current context of intensified global climate change, changes in meteorological elements, such as precipitation, temperature, and evaporation, have an important impact on basin hydrology and hydropower generation. Accurate prediction of runoff processes is crucial to the optimal scheduling of hydropower stations. The accuracy of the prediction determines whether the scheduling strategy can be effectively implemented, which in turn affects the economic benefits and operating efficiency of the hydropower station. Therefore, strengthening the research on runoff prediction under the background of climate change is of great significance to improving the operation and management level of hydropower stations, enhancing water resources regulation capabilities, and promoting the sustainable development of the hydropower industry.

[0059] Therefore, we need to solve the following problems to realize the prediction of power generation of run-of-river hydropower stations. First, at the data level, integrate multi-source hydrological data, including hydrological and meteorological data, satellite remote sensing data, ground monitoring network data, and numerical weather forecast data. Fully reflect the natural conditions of the basin and improve the prediction accuracy. Secondly, at the feature level, design effective feature extraction and fusion strategies, extract the potential features of various types of data through deep learning networks, and capture the inherent correlation between data. Finally, at the decision-making level, build an adaptive fusion decision framework. It is necessary to automatically adjust the contribution of each model to improve the robustness and accuracy of the prediction.

[0060] See also Figure 1 andFigure 2 , Figure 1 is the flowchart of the runoff hydropower station power generation prediction method based on deep data fusion; Figure 2 is the flow block diagram of the runoff hydropower station power generation prediction method based on deep data fusion.

[0061] This application proposes a runoff hydropower station power generation prediction method based on deep data fusion, including the following steps:

[0062] Step S1, data layer fusion stage. Integrate multi-source hydrometeorological data, obtain multi-source data from meteorological stations, hydrological stations, satellite remote sensing platforms and numerical weather prediction systems, and perform data cleaning, interpolation, normalization processing, and spatio-temporal alignment. Specifically include:

[0063] Step S11, data spatio-temporal alignment.

[0064] Step S111, input satellite remote sensing data, including precipitation, snow cover, and land cover, with a spatial resolution of 1km×1km; ground monitoring data, including rainfall, flow, and temperature, discrete site data; numerical weather prediction data, including precipitation, temperature, and grid data. Align the satellite remote sensing data and ground monitoring data spatially through the inverse distance weighting interpolation IDW method. The formula is:

[0065]

[0066] where Z(s0) is the target grid value, d(s0,s i ) is the distance from the site to the grid center, and p is the power parameter (usually taken as 2).

[0067] Step S112, perform time alignment through the sliding window aggregation method, and aggregate the satellite remote sensing data to the same granularity as the ground monitoring data (daily). Input the original time series data {(t i ,v i )} and the window parameter window size T (24 hours). For each target time point t' j , define the time window range as [t' j -T,t' j , extract all the original data points {v i} within the window, satisfying t' j -T≤t i ≤t' j , calculate the aggregation value Output the aligned time series

[0068] Step S12, introduce the generative adversarial network GAN to enhance the historical hydrometeorological data, generate synthetic data to alleviate the small sample problem, and output a multi-source data set with a unified spatio-temporal benchmark.

[0069] Construct a GAN model consisting of a generator and a discriminator. The principle is as follows:

[0070] The training process of GAN involves optimizing an adversarial loss function, which is usually expressed as a minimax game between the generator and the discriminator. The generator takes a random noise vector z ∼ p noise (z) as input and generates synthetic hydro-meteorological data G(z) through a multi-layer neural network; the discriminator classifies the real data x ∼ p data (x) and the generated data G(z) and outputs their authenticity probabilities. The adversarial loss function is:

[0071]

[0072] where D(x) is the output of the discriminator for the real data, and D(G(z)) is the output of the discriminator for the generated data.

[0073] During the training process, the Adam optimizer is used to update the network parameters, and its parameter update rule is:

[0074]

[0075] where and are the first and second moment estimates of the gradient respectively, and η is the learning rate. After training, the synthetic data output by the generator will be integrated with the real data into a unified multi-source dataset.

[0076] Step S2, Feature layer fusion stage. First, input the multi-source dataset with a unified spatio-temporal reference, perform feature extraction, and perform feature fusion based on the self-attention mechanism. And embed the water balance equation in the deep learning model, specifically including:

[0077] Step S21, Input the multi-source dataset with a unified spatio-temporal reference, extract the spatial features of satellite remote sensing data through 3D-CNN, extract the temporal features of ground time series data through LSTM, and extract the meteorological trend features of numerical weather prediction data through Transformer. Specifically including:

[0078] Step S211, Input satellite remote sensing data (spatial features) where H×W is the spatial resolution. D is the number of channels. T is the time step. Extract the spatial feature tensor through multi-layer convolution and pooling where T' is the time step, h×w is the spatial dimension, d is the number of feature channels, which is determined by the number of convolutional kernels;

[0079] Step S212, Input ground time series data (temporal features) Where T is the time step and F is the feature dimension. The time series dependencies such as traffic and temperature are captured by the LSTM unit to obtain the time series feature vector

[0080] Step S213, input numerical weather prediction data (meteorological trend features): Where T is the time step and C is the number of meteorological variables. The long-range dependencies of meteorological variables are captured by the multi-head attention mechanism to obtain the meteorological trend feature vector

[0081] Step S22, dynamically fuse features based on the attention mechanism, specifically including:

[0082] Step S221, calculate the attention score. Use the self-attention mechanism to calculate the mutual influence between features. First, project each modality feature H i into the query, key, and value spaces:

[0083] Q i = W Q H i , K i = W K H i , V i = W V H i

[0084] Where W Q , W K , W V are learnable projection matrices. Calculate the attention score Where d is the scaling factor of the feature dimension.

[0085] Normalize the attention scores

[0086] Step S222, calculate the weighted features. Use the attention weight α ij to perform weighted summation on the value vector V j :

[0087]

[0088] Step S23, embed the water balance equation. It includes:

[0089] Step S231, input the fused feature vector obtained in Step S222 and perform initial flow prediction through a fully connected network

[0090] Step S232, define the water balance equation:

[0091]

[0092] Among them, is the inflow, and ΔS (t) is the change in water storage, and the physical correction term ∈ is dynamically learned through the attention mechanism (t) . The predicted value of the outflow:

[0093]

[0094] The flow prediction result can be used as the ground monitoring data input in step S1 to improve the accuracy of the predicted hydropower generation.

[0095] Step S3, prediction stage. The DNN model architecture and training specifically include:

[0096] Step S31, forward propagation, calculate the hidden layer H i and the output layer

[0097] The input layer inputs the fused feature H fused , with dimension d. The hidden layer includes FC, the activation function ReLU, BN to prevent gradient vanishing, and the Dropout layer to prevent overfitting. The i-th hidden layer

[0098] H i = σ(W i H i-1 + b i )

[0099] Among them, H i-1 is the output of the previous layer, and the H obtained in step S222 of the first layer input fused , W i and b i are trainable parameters, and σ(·) is the activation function ReLU

[0100] The output layer performs a linear transformation to output the predicted value

[0101]

[0102] Among them, W out , b out are the output layer parameters, and H n is the output of the last hidden layer.

[0103] Step S32, train the DNN. Calculate the loss function loss

[0104]

[0105] Among them, Y i is the actual power generation, is the power generation predicted by the DNN.

[0106] The Adam function is used as the optimizer for training, and the optimization objective is to update W and b to minimize the loss.

[0107] In summary, this application proposes a runoff hydropower station power generation prediction method based on deep data fusion. First, multi-source data such as satellite remote sensing, ground monitoring, and numerical weather prediction are integrated, and the data is enhanced through spatio-temporal alignment (IDW interpolation, sliding window aggregation method) and GAN generative adversarial network to solve the problems of spatio-temporal inconsistency and small samples. Secondly, 3D-CNN, LSTM, and Transformer are used to extract spatial, temporal, and meteorological trend features respectively, the features are dynamically fused through the self-attention mechanism, and physical constraints are embedded in combination with the water balance equation. Finally, a DNN model is constructed for power generation prediction to achieve high-precision and interpretable runoff power generation prediction.

[0108] The applicant of the present invention has made a detailed description and illustration of the embodiments of the present invention in conjunction with the accompanying drawings of the specification. However, those skilled in the art should understand that the above embodiments are only the preferred implementation schemes of the present invention, and the detailed description is only to help readers better understand the spirit of the present invention, rather than a limitation on the protection scope of the present invention. On the contrary, any improvement or modification made based on the spirit of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for predicting power generation of a run-of-river hydropower station based on deep data fusion, characterized in that: The steps include: Step S1, perform data layer fusion, integrate multi-source hydrological and meteorological data, obtain multi-source data from meteorological stations, hydrological stations, satellite remote sensing platforms and numerical weather forecasting systems, and clean, interpolate and normalize the data to perform data spatiotemporal alignment; Step S2, feature layer fusion, first input a multi-source data set with a unified spatiotemporal reference, perform feature extraction, perform feature fusion based on the self-attention mechanism, embed the water balance equation in the deep learning model, and obtain a comprehensive feature vector; Step S3, making a prediction, inputting the comprehensive feature vector in step S2, and using a deep neural network to make a prediction of power generation.

2. The method for predicting power generation of a run-of-river hydropower station based on deep data fusion according to claim 1 is characterized in that: The step S1 specifically includes: Step S11, data spatiotemporal alignment: Step S111, spatially aligning satellite remote sensing data with ground monitoring data by using the inverse distance weighted method: Among them, Z(s0) is the target grid value, d(s0,s i ) is the distance from the site to the grid center, p is the power parameter; Step S112, time alignment is performed by sliding window aggregation method, the satellite remote sensing data is aggregated to a granularity consistent with the daily ground monitoring data, and the original time series data {(t i ,v i )} and window parameter window size T, for each target time point t j ' , define the time window range as [t' j -T,t' j ], extract all original data points in the window {v i }, satisfying t' j -T≤t i ≤t' j , calculate the aggregate value Output aligned time series Step S12, training a GAN model consisting of a generator and a discriminator, inputting the acquired historical hydrological and meteorological data, generating synthetic data by the generator, and outputting a multi-source data set with a unified spatiotemporal benchmark: Among them, D(x) is the output of the discriminator for real data, and D(G(z)) is the output of the discriminator for generated data.

3. The method for predicting power generation of a run-of-river hydropower station based on deep data fusion according to claim 2 is characterized in that: The step S2 specifically includes: Step S21, input a multi-source data set with a unified spatiotemporal reference, extract the spatial features of satellite remote sensing data through 3D-CNN, extract the temporal features of ground time series data through LSTM, and extract the meteorological trend features of numerical weather forecast data through Transformer, specifically including: Step S211, input satellite remote sensing data It belongs to the spatial feature, where H×W is the spatial resolution, D is the number of channels, and T is the time step. The spatial feature tensor is extracted through multi-layer convolution and pooling. Where T' is the time step, h×w is the spatial dimension, and d is the number of feature channels; Step S212: Input ground time series data It belongs to the time series feature, where T is the time step and F is the feature dimension. The time series dependencies such as flow and temperature are captured by LSTM units to obtain the time series feature vector Step S213, input numerical weather forecast data It belongs to the meteorological trend feature, where T is the time step, C is the number of meteorological variables, and the long-range dependency of meteorological variables is captured through the multi-head attention mechanism to obtain the meteorological trend feature vector Step S22, dynamically fusing features based on the attention mechanism, specifically includes: Step S221, calculate the attention score, and use the self-attention mechanism to calculate the mutual influence between features; first, each modal feature H i Projection into query, key and value spaces: Q i =W Q H i ,K i =W K H i ,V i =W V H i Among them, W Q ,W K ,W V is the learnable projection matrix; Calculating attention score Where d is the scaling factor of the feature dimension; Normalized attention score Step S222, calculate the fusion feature H fused , using the attention weight α ij The value vector V j Perform weighted summation: H fused =∑ j a ij V j ; Step S23, embedding the water balance equation, includes: Step S231: input the fusion feature vector obtained in step S222 and perform initial flow prediction through the fully connected network. Step S232, define the water balance equation: in, is the inflow rate, ΔS (t) is the change in water storage, and the physical correction term ∈ is dynamically learned through the attention mechanism (t) , output traffic prediction value: The water balance equation constraint ensures that the predicted value conforms to the physical laws and avoids the risk of overfitting driven by pure data. The flow prediction result can be used as the ground monitoring data input in step S1 to improve the accuracy of the power generation prediction of the hydropower station.

4. The method for predicting power generation of a run-of-river hydropower station based on deep data fusion according to claim 3 is characterized in that: In step S3, the DNN model architecture and training specifically include: Step S31, establish the DNN model framework, forward propagate, and calculate the hidden layer H i and output layer The input layer inputs the fused feature vector H obtained in step S222 fused , dimension d, hidden layers include multiple fully connected layers, activation function ReLU, batch normalization Dropout layer, the i-th hidden layer: H i =σ(W i H i-1 +b i ) Among them, H i-1 is the output of the previous layer, and the first layer input is H fused , W i ,b i is a trainable parameter, σ(·) is the activation function ReLU; The output layer linearly transforms the output prediction value: Among them, W out ,b out is the output layer parameter, H n is the output of the last hidden layer; Step S32, train the DNN model and calculate the loss function loss: Among them, Y i is the actual power generation, is the power generation predicted by the DNN model; The Adam function is used as the optimizer for training, and the optimization goal is to update W,b to minimize the loss.

Images