Water-wind-light multi-energy complementary output prediction method

Abstract

The invention discloses a water-wind-light multi-energy complementary output prediction method. The method comprises the steps of data collection and preprocessing, historical data collection, data preprocessing and sequence construction. Performing water, wind and light multi-source heterogeneous data fusion through the collected data, and integrating features from different energy types; performing post-fusion processing on the data subjected to the multi-source heterogeneous data fusion; executing model training and optimization; executing model evaluation and adjustment; according to the method, the output can be quickly predicted through a data fusion and analysis method after data acquisition, the calculation method can be applied to an edge end, and the pressure of the rear end is reduced; the problems that in the prior art, data of different data sources need to be uploaded, the mode of summation is mostly adopted after output is predicted separately, influences and relations among multi-source data are not considered, on the other hand, the data need to be uploaded to a far-end server to be analyzed, and efficiency is low are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-source heterogeneous data fusion, and in particular relates to a method for predicting the complementary output of water, wind and solar power. Background Art

[0002] The current output prediction in the field of water, wind and solar energy complementary energy is mainly to obtain and utilize historical output data sets, and then transmit the data to the server for analysis and prediction; the existing analysis and prediction methods need to upload data from different data sources, most of which adopt the method of summing up after predicting the output separately, without considering the influence and relationship between multi-source data. On the other hand, the data needs to be uploaded to a remote server for analysis, which is inefficient; therefore, it is necessary to design a water, wind and solar energy complementary energy output prediction method to solve the above problems. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method for predicting the complementary output of water, wind and solar power. This method can quickly predict the output through data fusion and analysis methods after data collection, and can be applied at the edge to reduce the pressure on the back end.

[0004] In order to achieve the above technical effects, the technical solution adopted by the present invention is: A method for predicting the complementary output of water, wind and solar power, comprising the following steps: Data collection and preprocessing: Collect historical data on the operation of hydropower, wind power and photovoltaic systems, perform data preprocessing and sequence construction on the collected historical data, and divide the preprocessed and sequence-constructed data into training sets and test sets according to a fixed ratio; Construct a Transformer-based water-wind-solar multi-energy complementary output prediction model, and train and optimize the water-wind-solar multi-energy complementary output prediction model through the training set; Perform multi-source heterogeneous data fusion on the collected historical data of hydropower, wind power and photovoltaic systems to integrate the features from different energy types; The data after fusion of multi-source heterogeneous data is post-fused and input into the water-wind-solar multi-energy complementary output prediction model for output prediction.

[0005] Preferably, the historical data includes historical output data of hydropower, wind power and photovoltaic systems; the historical output data includes operation characteristic data, meteorological characteristic data and time characteristic data; the operation characteristic data includes water level height, water flow velocity, wind speed, wind direction, light intensity, power demand, weather conditions, time, maintenance records, historical energy output and consumption data; the meteorological characteristic data includes temperature, humidity, wind speed and light intensity, and the time characteristic data includes hours and seasons.

[0006] Preferably, data preprocessing includes: Data cleaning: clean the historical data of hydropower, wind power and photovoltaic system operation and process missing values; Data standardization: standardize or normalize numerical data and convert categorical data into numerical codes; Time series data alignment: Align historical data on the operation of hydropower, wind power and photovoltaic systems from different sources on the same time scale.

[0007] Preferably, the sequence construction includes: The data is organized into a time series format, where each data sample contains the data points of the past T time period and the output corresponding to the future time point t0 as a label.

[0008] Preferably, a Transformer-based water-wind-solar multi-energy complementary output prediction model is established, and the training and optimization of the water-wind-solar multi-energy complementary output prediction model includes: A Transformer-based water, wind and solar power complementary output prediction model is constructed. The model includes an encoder layer, an embedding layer, a fully connected layer and a linear layer. The encoder layer includes multiple sub-layers including a multi-head self-attention layer, a feedforward neural network layer, a residual connection layer and a layer normalization layer. Minimize loss: During the training process of the hydro-wind-solar multi-energy complementary output prediction model, all parameters, including those of the modal feature extractor and the linear layer, are adjusted through back propagation to minimize the prediction error or loss function; Dynamically adjust the learning rate: Regularization techniques are used to prevent overfitting, and learning rate strategies are dynamically adjusted to accelerate convergence and improve performance. Use the cross-validation method to evaluate the generalization ability of the model, and adjust the fusion strategy or the entire model structure based on the evaluation results.

[0009] Preferably, the collected historical data of hydropower, wind power and photovoltaic systems are fused with multi-source heterogeneous data to integrate the features from different energy types, including: Modal alignment: The data of three modalities, hydropower, wind power and photovoltaic power, are mapped into a common low-dimensional space through projection alignment.

[0010] Preferably, the collected historical data of hydropower, wind power and photovoltaic systems are fused with multi-source heterogeneous data to integrate the features from different energy types, including: The embedding layer of the hydro-wind-solar multi-energy complementary output prediction model extracts the time series periodic features of different types of variables such as hours and seasons, operation features, meteorological features, and time features, and sends them to the encoder layer of the Transformer for feature extraction. Finally, the extracted features are fused and sent to the linear layer for residual prediction. Multi-head attention mechanism processing: The attention mechanism is applied separately to the data of the three modes of hydropower, wind power and photovoltaic power to extract key information.

[0011] Preferably, the collected historical data of hydropower, wind power and photovoltaic systems are fused with multi-source heterogeneous data to integrate the features from different energy types, including: Attention score calculation: For each mode or feature within a mode in the three modal data of hydropower, wind power and photovoltaic, the attention score is calculated through the interaction of Query, Key and Value: Query represents the information that currently needs attention, Key is the information source, and Value is the actual information content selected based on the Key.

[0012] Preferably, the collected historical data of hydropower, wind power and photovoltaic systems are fused with multi-source heterogeneous data to integrate the features from different energy types, including: Weighted summation: Using the attention score as the weight, the values ​​of the three modal data of hydropower, wind power and photovoltaic are weightedly summed to form a fused feature vector representation.

[0013] Preferably, post-fusion processing is performed on the data after fusion of multi-source heterogeneous data, and the post-fusion processed data is input into the hydro-wind-solar multi-energy complementary output prediction model for output prediction, which includes: The weighted summed feature vector is sent to the linear layer for calculation, and the output is predicted using linear regression; Determine the loss function and optimization algorithm: The mean square error MSE is used as the loss function for training; Choose Adam or RMSprop optimization algorithm to minimize the loss function.

[0014] The beneficial effects of the present invention are as follows: The present invention proposes a method for predicting the complementary output of water, wind and solar power, so that the output can be quickly predicted through data fusion and analysis methods after data collection. The calculation method can be applied at the edge to reduce the pressure on the back end; it solves the problem that the prior art needs to upload data from different data sources, and most of them adopt the method of summing up after separately predicting the output, without considering the influence and relationship between multi-source data. On the other hand, the data needs to be uploaded to a remote server for analysis, which is inefficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The figure is a flowchart of the method of the present invention. DETAILED DESCRIPTION

[0016] Embodiment 1: like Figure 1 As shown, a method for predicting the complementary output of water, wind and solar power includes the following steps: Data collection and preprocessing: Collect historical data on the operation of hydropower, wind power and photovoltaic systems, perform data preprocessing and sequence construction on the collected historical data, and divide the preprocessed and sequence-constructed data into training sets and test sets according to a fixed ratio; Construct a Transformer-based water-wind-solar multi-energy complementary output prediction model, and train and optimize the water-wind-solar multi-energy complementary output prediction model through the training set; Perform multi-source heterogeneous data fusion on the collected historical data of hydropower, wind power and photovoltaic systems to integrate the features from different energy types; The data after fusion of multi-source heterogeneous data is post-fused and input into the water-wind-solar multi-energy complementary output prediction model for output prediction.

[0017] Preferably, the historical data includes historical output data of hydropower, wind power and photovoltaic systems; the historical output data includes operation characteristic data, meteorological characteristic data and time characteristic data; the operation characteristic data includes water level height, water flow velocity, wind speed, wind direction, light intensity, power demand, weather conditions, time, maintenance records, historical energy output and consumption data; the meteorological characteristic data includes temperature, humidity, wind speed and light intensity, and the time characteristic data includes hours and seasons.

[0018] Preferably, data preprocessing includes: Data cleaning: clean the historical data of hydropower, wind power and photovoltaic system operation and process missing values; Data standardization: standardize or normalize numerical data and convert categorical data into numerical codes; Time series data alignment: Align historical data on the operation of hydropower, wind power and photovoltaic systems from different sources on the same time scale.

[0019] Preferably, the sequence construction includes: The data is organized into a time series format, where each data sample contains the data points of the past T time period and the output corresponding to the future time point t0 as a label.

[0020] Preferably, a Transformer-based water-wind-solar multi-energy complementary output prediction model is established, and the training and optimization of the water-wind-solar multi-energy complementary output prediction model includes: A Transformer-based water, wind and solar power complementary output prediction model is constructed. The model includes an encoder layer, an embedding layer, a fully connected layer and a linear layer. The encoder layer includes multiple sub-layers including a multi-head self-attention layer, a feedforward neural network layer, a residual connection layer and a layer normalization layer. Minimize loss: During the training process of the hydro-wind-solar multi-energy complementary output prediction model, all parameters, including those of the modal feature extractor and the linear layer, are adjusted through back propagation to minimize the prediction error or loss function; Dynamically adjust the learning rate: Regularization techniques are used to prevent overfitting, and learning rate strategies are dynamically adjusted to accelerate convergence and improve performance. Use the cross-validation method to evaluate the generalization ability of the model, and adjust the fusion strategy or the entire model structure based on the evaluation results.

[0021] Preferably, the collected historical data of hydropower, wind power and photovoltaic systems are fused with multi-source heterogeneous data to integrate the features from different energy types, including: Modal alignment: The data of three modalities, hydropower, wind power and photovoltaic power, are mapped into a common low-dimensional space through projection alignment.

[0022] Preferably, the collected historical data of hydropower, wind power and photovoltaic systems are fused with multi-source heterogeneous data to integrate the features from different energy types, including: The embedding layer of the hydro-wind-solar multi-energy complementary output prediction model extracts the time series periodic features of different types of variables such as hours and seasons, operation features, meteorological features, and time features, and sends them to the encoder layer of the Transformer for feature extraction. Finally, the extracted features are fused and sent to the linear layer for residual prediction. Multi-head attention mechanism processing: The attention mechanism is applied separately to the data of the three modes of hydropower, wind power and photovoltaic power to extract key information.

[0023] Preferably, the collected historical data of hydropower, wind power and photovoltaic systems are fused with multi-source heterogeneous data to integrate the features from different energy types, including: Attention score calculation: For each mode or feature within a mode in the three modal data of hydropower, wind power and photovoltaic, the attention score is calculated through the interaction of Query, Key and Value: Query represents the information that currently needs attention, Key is the information source, and Value is the actual information content selected based on the Key.

[0024] Preferably, the collected historical data of hydropower, wind power and photovoltaic systems are fused with multi-source heterogeneous data to integrate the features from different energy types, including: Weighted summation: Using the attention score as the weight, the values ​​of the three modal data of hydropower, wind power and photovoltaic are weightedly summed to form a fused feature vector representation.

[0025] Preferably, post-fusion processing is performed on the data after fusion of multi-source heterogeneous data, and the post-fusion processed data is input into the hydro-wind-solar multi-energy complementary output prediction model for output prediction, which includes: The weighted summed feature vector is sent to the linear layer for calculation, and the output is predicted using linear regression; Determine the loss function and optimization algorithm: The mean square error MSE is used as the loss function for training; Choose Adam or RMSprop optimization algorithm to minimize the loss function.

[0026] Embodiment 2: Preferably, data preprocessing includes: Data cleaning: Clean the data of each modality and handle missing values; Furthermore, for fields with missing values, linear interpolation is used to fill in the missing values. The interpolation polynomial is as shown below: ; In the formula, y 0 and y 1 is the two values ​​at the time around the abnormal point; linear interpolation is performed through the missing value t The coordinates of the time ( x,y ), and the coordinate points of the two moments to the left and right t -1 moment( x 0 ,y 0), No. t +1( x 1 ,y 1) and x Worth the moment y The value is estimated, Indicates t Other non-missing eigenvalues ​​at time.

[0027] Data standardization: standardize or normalize numerical data and convert categorical data into numerical codes; Preferably, the data is min-max normalized: ; In the formula, x max In the sample data x The maximum value of x min In the sample datax The minimum value of .

[0028] For categorical features in the data, one-hot encoding is used for discretization encoding processing to convert categorical data into numerical encoding.

[0029] Discrete variables such as season type and week type are encoded using sine-cosine cyclic encoder: ; In the formula, x represents the index number, x max Represents the maximum value of the index number; When encoding the season type, x∈{1, 2, 3}; when encoding the hour type, x∈{1, .....,24}; for the category type, one-hot encoding is used, and different categories are encoded as 0 or 1.

[0030] Time series data alignment: Align data from different sources on the same time scale; aggregate data at different times according to a unified time to achieve data alignment.

[0031] Preferably, the sequence construction includes: The data is organized into a time series format, where each data sample contains the data points of the past T time period and the output at the corresponding future time point t0 as a label; for example, if the output for the next day is predicted, each sample may contain the daily output data of hydropower, wind power, and solar power for the previous 30 days.

[0032] Preferably, the water, wind and solar multi-source heterogeneous data fusion through the collected data includes: Modal alignment: Since the feature spaces of different modalities may be inconsistent, they need to be mapped to a shared space or a common representation form to facilitate fusion; the data of the three modalities of water, wind and light are mapped to a common low-dimensional space through projection alignment, so that the representations of the three modalities in this space maintain relevance to the greatest extent, which is convenient for subsequent prediction calculations.

[0033] Furthermore, the data of the three modes of water, wind and light are mapped into a common low-dimensional space through projection alignment. The specific method is to construct a calculation formula for feature projection according to the category, operation characteristics, meteorological characteristics and time characteristics of the data: Category, including the type of hydropower, wind power or photovoltaic data belongs to, defined as x a ; Operational characteristics, including water level, water velocity, wind speed, wind direction, light intensity, power demand, weather conditions, time of day, maintenance records, historical energy production and consumption data, are defined as xb ; Meteorological characteristic data, including temperature, humidity, wind speed and light intensity, are defined as x c ; Temporal feature data includes hours and seasons, defined as x d ; Define a nonlinear neural network mapping layer to project these features into the same low-dimensional mapping space. The neural network mapping layer can be the fully connected layer in the constructed hydro-wind-solar multi-energy complementary output prediction model. Suppose the weight matrix of the mapping layer is W, the bias is B, the activation function is f, and the projection process is: ; Where z is the result of projection to the same low-dimensional mapping space.

[0034] Preferably, the water, wind and solar multi-source heterogeneous data fusion through the collected data includes: Multi-head attention mechanism processing: Apply attention mechanism to each type of data of each modality separately to extract key information; The self-attention mechanism allows the model to consider the dependencies between different positions within the sequence when calculating the hidden representation. The self-attention mechanism can adaptively assign different weights to different positions, thereby capturing information within the sequence. Its calculation formula is as follows: ; in, Denotes the self-attention weight calculation, the superscript T is the matrix transpose, is the dimension of the key; softmax represents the linear activation function; each head focuses on a different feature subspace, thereby capturing diverse information.

[0035] Preferably, the water, wind and solar multi-source heterogeneous data fusion through the collected data includes: Attention score calculation: For each modality or feature within a modality, the attention score is calculated by the interaction of Q, K and V: The calculation process is: ; Q represents the information that needs attention at present, K represents i represents the i-th information source, m represents the number of information sources, and V is the actual information content selected according to Key; for example, for the predicted output of wind energy, Q represents the problem of wind energy output, k represents the label of wind speed, and V represents the value of actual wind speed.

[0036] Preferably, the water, wind and solar multi-source heterogeneous data fusion through the collected data includes: Weighted sum: Using the attention score as the weight, the values ​​of different modes are weighted and summed to form the fused feature vector representation: ; Among them, among them, represents the query matrix, Represents a repository of data from different types of sources, represents the value matrix, represents the attention score of the i-th information source, and Lx represents the information library Length.

[0037] Preferably, the post-fusion processing includes: The weighted summed feature vector is sent to the linear layer for calculation, and the output is predicted using linear regression; Determine the loss function and optimization algorithm: The mean square error MSE is used as the loss function for training; ; In the formula, m represents the number of calibration samples; Indicates i The actual value of samples; Indicates i The predicted value of a sample.

[0038] Choose Adam or RMSprop optimization algorithm to minimize the loss function.

[0039] Preferably, model training and optimization include: Minimize loss: During the entire network training process, all parameters, including those of the modality feature extractor and fusion layer, are adjusted via back-propagation to minimize the prediction error or loss function.

[0040] Dynamically adjust the learning rate: Regularization techniques are used to prevent overfitting, and learning rate strategies are dynamically adjusted to accelerate convergence and improve performance. Use the cross-validation method to evaluate the generalization ability of the model, and adjust the fusion strategy or the entire model structure based on the evaluation results.

Claims

1. A method for predicting the complementary output of water, wind and solar energy, characterized in that: The following steps are involved: Data collection and preprocessing: Collect historical data on the operation of hydropower, wind power and photovoltaic systems, perform data preprocessing and sequence construction on the collected historical data, and divide the preprocessed and sequence-constructed data into training sets and test sets according to a fixed ratio; Construct a Transformer-based water-wind-solar multi-energy complementary output prediction model, and train and optimize the water-wind-solar multi-energy complementary output prediction model through the training set; Perform multi-source heterogeneous data fusion on the collected historical data of hydropower, wind power and photovoltaic systems to integrate the features from different energy types; The data after fusion of multi-source heterogeneous data is post-fused and input into the water-wind-solar multi-energy complementary output prediction model for output prediction.

2. The method for predicting the complementary output of water, wind and solar energy according to claim 1 is characterized in that: Historical data include historical output data of hydropower, wind power and photovoltaic systems; historical output data include operation characteristic data, meteorological characteristic data and time characteristic data; Operational characteristic data include water level, water flow velocity, wind speed, wind direction, light intensity, power demand, weather conditions, time, maintenance records, historical energy output and consumption data; meteorological characteristic data include temperature, humidity, wind speed and light intensity; time characteristic data include hours and seasons.

3. The method for predicting the complementary output of water, wind and solar energy according to claim 2 is characterized in that: Data preprocessing includes: Data cleaning: clean the historical data of hydropower, wind power and photovoltaic system operation and process missing values; Data standardization: standardize or normalize numerical data and convert categorical data into numerical codes; Time series data alignment: Align historical data on the operation of hydropower, wind power and photovoltaic systems from different sources on the same time scale.

4. The method for predicting the complementary output of water, wind and solar energy according to claim 3 is characterized in that: The sequence construction includes: The data is organized into a time series format, where each data sample contains the data points of the past T time period and the output corresponding to the future time point t0 as a label.

5. The method for predicting the complementary output of water, wind and solar energy according to claim 4 is characterized in that: Establish a Transformer-based hydro-wind-solar multi-energy complementary output prediction model, and train and optimize the hydro-wind-solar multi-energy complementary output prediction model, including: A Transformer-based water, wind and solar power complementary output prediction model is constructed. The model includes an encoder layer, an embedding layer, a fully connected layer and a linear layer. The encoder layer includes multiple sub-layers including a multi-head self-attention layer, a feedforward neural network layer, a residual connection layer and a layer normalization layer. Minimize loss: During the training process of the hydro-wind-solar multi-energy complementary output prediction model, all parameters, including those of the modal feature extractor and the linear layer, are adjusted through back propagation to minimize the prediction error or loss function; Dynamically adjust the learning rate: Regularization techniques are used to prevent overfitting, and learning rate strategies are dynamically adjusted to accelerate convergence and improve performance. Use the cross-validation method to evaluate the generalization ability of the model, and adjust the fusion strategy or the entire model structure based on the evaluation results.

6. The method for predicting the complementary output of water, wind and solar energy according to claim 5 is characterized in that: The collected historical data of hydropower, wind power and photovoltaic systems are integrated with multi-source heterogeneous data to integrate the features from different energy types, including: Modal alignment: The data of three modalities, hydropower, wind power and photovoltaic power, are mapped into a common low-dimensional space through projection alignment.

7. The method for predicting the complementary output of water, wind and solar energy according to claim 6 is characterized in that: The collected historical data of hydropower, wind power and photovoltaic systems are integrated with multi-source heterogeneous data to integrate the features from different energy types, including: The embedding layer of the hydro-wind-solar multi-energy complementary output prediction model extracts the time series periodic features of different types of variables such as hours and seasons, operation features, meteorological features, and time features, and sends them to the encoder layer of the Transformer for feature extraction. Finally, the extracted features are fused and sent to the linear layer for residual prediction. Multi-head attention mechanism processing: The attention mechanism is applied separately to the data of the three modes of hydropower, wind power and photovoltaic power to extract key information.

8. The method for predicting the complementary output of water, wind and solar energy according to claim 7 is characterized in that: The collected historical data of hydropower, wind power and photovoltaic systems are integrated with multi-source heterogeneous data to integrate the features from different energy types, including: Attention score calculation: For each mode or feature within a mode in the three modal data of hydropower, wind power and photovoltaic, the attention score is calculated through the interaction of Query, Key and Value: Query represents the information that currently needs attention, Key is the information source, and Value is the actual information content selected based on the Key.

9. The method for predicting the complementary output of water, wind and solar energy according to claim 8 is characterized in that: The collected historical data of hydropower, wind power and photovoltaic systems are integrated with multi-source heterogeneous data to integrate the features from different energy types, including: Weighted summation: Using the attention score as the weight, the values ​​of the three modal data of hydropower, wind power and photovoltaic are weightedly summed to form a fused feature vector representation.

10. The method for predicting the complementary output of water, wind and solar energy according to claim 9 is characterized in that: The data after fusion of multi-source heterogeneous data is post-fused and input into the hydro-wind-solar multi-energy complementary output prediction model for output prediction, including: The weighted summed feature vector is sent to the linear layer for calculation, and the output is predicted using linear regression; Determine the loss function and optimization algorithm: The mean square error MSE is used as the loss function for training; Choose Adam or RMSprop optimization algorithm to minimize the loss function.

Images