A power prediction method based on probability space transformation and neural network fusion
By fusing probability space transformation with neural networks, the problem of capturing complex nonlinear relationships in new energy power generation prediction has been solved, achieving high-precision and stable wind power prediction, which is applicable to complex atmospheric environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- POWERCHINA HUADONG ENG CORP LTD
- Filing Date
- 2026-01-16
- Publication Date
- 2026-06-05
AI Technical Summary
Existing methods for predicting new energy power generation struggle to accurately capture the complex nonlinear relationships between multiple variables when faced with the intermittency and uncertainty of natural resources such as wind and solar energy, resulting in insufficient prediction accuracy and stability.
We employ a method that combines probability space transformation with neural networks. By constructing a one-dimensional probability analyzer, we map and transform feature data. We combine multi-layer stacking and residual connection mechanisms to improve feature extraction capabilities and output prediction results through a fully connected regression network.
It significantly improves the accuracy and stability of wind power prediction, effectively addresses the problems of non-Gaussian and multi-peak distributions, adapts to complex atmospheric environments, and has good prospects for engineering applications.
Smart Images
Figure CN122159175A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of interdisciplinary application technology of artificial intelligence and power engineering, specifically involving a power prediction method based on the fusion of probability space transformation and neural network. Background Technology
[0002] As renewable energy accounts for an increasing proportion of the energy mix, forecasting renewable energy power generation is playing an increasingly important role in grid dispatch, energy management, and electricity market operation. However, due to the intermittency and uncertainty of natural resources such as wind and solar energy, their corresponding power generation exhibits significant volatility and non-stationarity, posing a considerable challenge to accurate forecasting.
[0003] Currently, methods for predicting new energy power mainly include physical models, statistical models, and machine learning-based prediction methods. Among them, physical models rely on the accuracy of meteorological simulations and have high computational costs; traditional statistical models are difficult to effectively capture the complex nonlinear relationships between multiple variables; and although deep learning models, which have been widely used in recent years, have improved their modeling capabilities, they often lack explicit modeling of the distribution characteristics of input features, resulting in limited generalization performance when faced with non-Gaussian or multimodal data.
[0004] Several power prediction methods based on deep neural networks have been proposed and applied in engineering practice. For example:
[0005] Chinese patent CN109242200B proposes a method for predicting wind power range using a Bayesian network prediction model. This method achieves probabilistic prediction by constructing Bayesian network relationships, but it suffers from computational efficiency bottlenecks when processing large-scale real-time data.
[0006] Chinese patent CN113256407A proposes a wind power prediction method based on LSTM and attention mechanism, which improves prediction accuracy by utilizing time series modeling capabilities, but does not consider the spatial distribution characteristics of input features. Summary of the Invention
[0007] The main objective of this invention is to provide a power prediction method based on the fusion of probability space transformation and neural network, addressing the aforementioned problems.
[0008] Therefore, the above-mentioned objective of the present invention is achieved through the following technical solution:
[0009] A power prediction method based on the fusion of probability space transformation and neural network includes the following steps:
[0010] S1. Construct a one-dimensional probability analyzer: Obtain the input feature data and construct a one-dimensional probability analyzer. The one-dimensional probability analyzer is used to establish a lookup table for the empirical cumulative distribution function (CDF) and the quantile function (PPF).
[0011] S2. Probability Space Mapping and Transformation: The input features are mapped to the probability space using a lookup table to obtain the corresponding cumulative probability value. The cumulative probability value is then transformed using a nonlinear neural network to output the transformed probability value. Finally, the transformed probability value is mapped back to the original feature space using a PPF lookup table to obtain the transformed indirect feature representation.
[0012] S3. Feature fusion mechanism: The original input features are processed by a linear neural network to extract direct features, which are then added to and fused with the obtained indirect features to obtain a fused feature representation.
[0013] S4. Multi-layer stacking and residual connection mechanism: The fused feature representation is input into multiple cascaded probability transformation modules to form a multi-layer stacked structure, and the residual connection mechanism is combined to perform layer-by-layer feature extraction.
[0014] S5. Output of Fully Connected Regression Network: The final power prediction result is output through the fully connected regression network.
[0015] S6. Model Training and Evaluation: Based on steps S2-S5, perform end-to-end training on the overall model, use the model with the smallest error in the validation set as the optimal training model, and test the accuracy of the model's predictions on the test set.
[0016] While adopting the above technical solutions, the present invention may also adopt or combine the following technical solutions:
[0017] As a preferred technical solution of the present invention: In step S1, the input feature data needs to be cleaned to remove outliers and to divide the training set, validation set and test set.
[0018] As a preferred technical solution of the present invention: In step S1, the one-dimensional probability analyzer receives input data of shape (B,L,C), where B represents the batch size, L represents the number of samples, and C represents the number of variables.
[0019] As a preferred technical solution of the present invention: in step S1, the empirical cumulative distribution function (CDF) lookup table calculates the cumulative probability corresponding to any input value by interpolation.
[0020] As a preferred technical solution of the present invention: in step S1, the quantile function PPF lookup table calculates the original value corresponding to any probability value based on the percentile.
[0021] As a preferred embodiment of the present invention, the nonlinear neural network consists of a linear transformation layer, a Sigmoid activation function layer, a linear transformation layer, and a Sigmoid activation function layer connected in sequence.
[0022] As a preferred technical solution of the present invention: in step S4, the residual connection mechanism is to add the output of the first layer probability transformation module to the output of each subsequent layer element by element.
[0023] As a preferred technical solution of the present invention: In step S5, the fully connected regression network includes at least one linear transformation layer, which combines the ReLU activation function and the Dropout regularization mechanism to finally output a single-value power prediction result.
[0024] As a preferred embodiment of the present invention: in step S6, the following analyzer generation step is further included in the training phase:
[0025] S61. Generate several slightly different one-dimensional probability analyzers based on the original training data;
[0026] S62. Differences are achieved by introducing perturbations into the original data, including but not limited to data scaling or adding Gaussian noise.
[0027] S63. Each probability analyzer corresponds to a probability transformation module of a different layer to improve model diversity and generalization ability.
[0028] Compared with the prior art, the present invention has the following beneficial effects: the method of the present invention can effectively deal with the non-Gaussian and multi-peak distribution of meteorological variables such as wind speed and wind direction, has strong adaptability to complex atmospheric environments, significantly improves the accuracy and stability of wind power prediction, and has good engineering application prospects. Attached Figure Description
[0029] Figure 1 The flowchart shows the power prediction method based on the fusion of probability space transformation and neural network provided by the present invention.
[0030] Figure 2 This is a schematic diagram of the probability change module architecture.
[0031] Figure 3 A flowchart for constructing a one-dimensional probability analyzer.
[0032] Figure 4 This is a comparison chart of predicted and actual operating power of wind farms. Detailed Implementation
[0033] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0034] like Figure 1 As shown, a power prediction method based on the fusion of probability space transformation and neural network includes the following steps:
[0035] S1. Construct a one-dimensional probabilistic analyzer: Obtain input feature data, including historical meteorological forecast data (such as wind speed, wind direction, temperature and air pressure at different altitudes) and historical power data of wind farms. Through data cleaning, outliers are removed, and training, validation and test sets are divided. For each training set input feature, construct a one-dimensional probabilistic analyzer.
[0036] like Figure 3 As shown, an empirical cumulative distribution function lookup table (CDF table) and a quantile function lookup table (PPF table) are constructed for each variable. The empirical cumulative distribution function CDF lookup table calculates the cumulative probability corresponding to any input value through interpolation, while the quantile function PPF lookup table calculates the original value corresponding to any probability value based on the percentile.
[0037] The number of probability analyzers is the number of probability transformation modules set in the model during training, and the statistical data is expressed by the following formula:
[0038] In the formula, the noise follows a random normal distribution, i.e. , where n is the number of probability change layers.
[0039] A one-dimensional probability analyzer receives input data of shape (B, L, C), where B represents the batch size, L represents the number of samples, and C represents the number of variables.
[0040] S2. Probability space mapping and transformation: such as Figure 2 As shown, the input features are divided into two paths.
[0041] Path 1: Map the input features to the probability space through the CDF lookup table in step S1 to obtain the corresponding cumulative probability value; after transformation by a nonlinear neural network (linear layer + Sigmoid + linear layer + Sigmoid), output a new probability value; then map it back to the original space through the PPF lookup table to obtain the indirect feature representation.
[0042] Path 2: Input features are directly extracted through a linear neural network.
[0043] S21. Use a lookup table to map the input features to the probability space to obtain the corresponding cumulative probability values;
[0044] S22. Transform the cumulative probability value using a nonlinear neural network and output the transformed probability value;
[0045] A nonlinear neural network consists of a linear transformation layer, a sigmoid activation function layer, another linear transformation layer, and another sigmoid activation function layer connected in sequence.
[0046] S23. Use the PPF lookup table to map the transformed probability values back to the original feature space to obtain the transformed indirect feature representation.
[0047] S3. Feature fusion mechanism: The original input features are processed by a linear neural network to extract direct features, which are then added to and fused with the indirect features obtained in step S23 to obtain a fused feature representation.
[0048] S4. Multi-layer stacking and residual connection mechanism: The fused feature representation is input into multiple cascaded probability transformation modules to form a multi-layer stacked structure, and the residual connection mechanism is combined to perform layer-by-layer feature extraction.
[0049] The residual connection mechanism involves adding the output of the first-layer probability transformation module element-wise with the output of each subsequent layer to enhance model stability.
[0050] S5. Fully Connected Regression Network Output: Input the multi-layer transformed feature representation into a fully connected regression network consisting of several linear layers (at least one linear transformation layer), and combine the ReLU activation function and Dropout regularization mechanism to output a single-value wind farm power prediction result.
[0051] S6. Model Training and Evaluation: Based on steps S2-S5, perform end-to-end training on the overall model, use the model with the smallest error in the validation set as the optimal training model, and test the accuracy of the model's predictions on the test set.
[0052] To improve model diversity and generalization performance during model training, the following analyzer generation steps are also included:
[0053] S61. Generate several slightly different one-dimensional probability analyzers based on the original training data;
[0054] S62. Differences are achieved by introducing perturbations into the original data, including but not limited to data scaling or adding Gaussian noise.
[0055] S63. Each probability analyzer corresponds to a probability transformation module of a different layer, thereby enhancing the model's adaptability to changes in input distribution and improving model diversity and generalization ability.
[0056] Example
[0057] Taking an onshore wind farm as the research object, this study explored the practical application effect of the method of this invention in wind power prediction. Two years of raw wind power data with a resolution of 15 minutes were collected, along with corresponding numerical meteorological simulation data. Input features included wind speed and direction at heights of 10 m, 100 m, and 200 m, and air temperature and pressure at 10 m, totaling eight features. The shape of the input features was (B, L, C), where the batch size B was set to 32, the data length L=24 (i.e., predicting 6 hours of wind power each time), and the number of channels C=8.
[0058] The model consists of three probability transformation layers and two fully connected layers. Mean squared error (MSE) is used as the loss function during training, with Adam as the optimizer and a learning rate of 1e-3. The training, validation, and test sets are partitioned in a 7:2:1 ratio. The model's prediction performance on the test set is as follows: Figure 4 As shown, the correlation coefficient with the actual operating data is as high as 85% or more, which meets the requirements for short-term power prediction.
[0059] Meanwhile, to verify the effectiveness of the model, two traditional neural networks were selected as comparison objects: linear neural networks and convolutional neural networks. Their corresponding statistical indicators were root mean square error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient (R²). 2 As shown in Table 1 below:
[0060] Table 1
[0061] Experimental results show that the method of the present invention outperforms traditional neural network models in multiple evaluation metrics such as RMSE, MAE, and R², and has higher prediction accuracy and stability.
[0062] The technical solution of the present invention has been described in conjunction with the specific experimental procedures shown in the accompanying drawings. However, the scope of protection of the present invention is not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions resulting from such changes or substitutions will all fall within the scope of protection of the present invention.
Claims
1. A power prediction method based on the fusion of probability space transformation and neural network, characterized in that, Includes the following steps: S1. Obtain the input feature data and construct a one-dimensional probability analyzer. The one-dimensional probability analyzer is used to establish the empirical cumulative distribution function (CDF) and quantile function (PPF) lookup table. S2. Use a lookup table to map the input features to the probability space to obtain the corresponding cumulative probability value. Transform the cumulative probability value through a nonlinear neural network and output the transformed probability value. Use a PPF lookup table to map the transformed probability value back to the original feature space to obtain the transformed indirect feature representation. S3. Extract direct features from the original input features using a linear neural network, and then add and fuse them with the obtained indirect features to obtain a fused feature representation; S4. Input the fused feature representation into multiple cascaded probability transformation modules to form a multi-layer stacked structure, and perform layer-by-layer feature extraction in conjunction with the residual connection mechanism. S5. Output the final power prediction result through a fully connected regression network; S6. Based on steps S2-S5, perform end-to-end training on the overall model, use the model with the smallest error in the validation set as the optimal training model, and test the accuracy of the model's predictions in the test set.
2. The method according to claim 1, characterized in that: In step S1, the input feature data needs to be cleaned to remove outliers and to be divided into training, validation and test sets.
3. The method according to claim 1, characterized in that: In step S1, the one-dimensional probability analyzer receives input data of shape (B,L,C), where B represents the batch size, L represents the number of samples, and C represents the number of variables.
4. The method according to claim 1, characterized in that: In step S1, the empirical cumulative distribution function (CDF) lookup table calculates the cumulative probability corresponding to any input value through interpolation.
5. The method according to claim 1, characterized in that: In step S1, the quantile function PPF lookup table calculates the original value corresponding to any probability value based on the percentile.
6. The method according to claim 1, characterized in that: A nonlinear neural network consists of a linear transformation layer, a sigmoid activation function layer, another linear transformation layer, and another sigmoid activation function layer connected in sequence.
7. The method according to claim 1, characterized in that: In step S4, the residual connection mechanism involves adding the output of the first-layer probability transformation module to the output of each subsequent layer element by element.
8. The method according to claim 1, characterized in that: In step S5, the fully connected regression network includes at least one linear transformation layer, which combines the ReLU activation function and the Dropout regularization mechanism to finally output a single-value power prediction result.
9. The method according to claim 1, characterized in that: Step S6, during the training phase, also includes the following analyzer generation step: S61. Generate several slightly different one-dimensional probability analyzers based on the original training data; S62. Differences are achieved by introducing perturbations into the original data, including but not limited to data scaling or adding Gaussian noise. S63. Each probability analyzer corresponds to a probability transformation module of a different layer to improve model diversity and generalization ability.