Renewable energy product prediction method based on deep learning

Abstract

The invention discloses a renewable energy product prediction method based on deep learning, and relates to the field of biomass technology and the like. The method comprises the following steps: firstly, acquiring a real data set, performing preprocessing such as cleaning and normalization, and screening parameters with relatively high correlation with a target variable; then, performing data enhancement by combining a variational auto-encoder and a generative adversarial network; constructing a support vector regression model, and optimizing model parameters by adopting a particle swarm optimization algorithm; and finally, predicting the performance through the index evaluation model. Compared with other methods, the method has the advantages that the problems of data scarcity and more raw material characteristics are effectively solved, the prediction precision, reliability and system robustness of the biomass co-pyrolysis bio-oil yield are improved, the method can fully play a role in the field of biomass energy, and support is provided for related scenes such as energy production and environment restoration.

Description

Technical Field

[0001] This invention relates to the field of biomass technology, specifically to a method for predicting renewable energy products based on deep learning. Background Technology

[0002] In recent years, with the global energy structure transitioning towards green and low-carbon development, renewable energy power generation technologies, represented by biomass energy, have received increasing attention. Biomass co-pyrolysis technology is gradually becoming an important research direction in the energy sector and environmental science, serving as a crucial upstream link in obtaining stable and controllable green electricity in biomass power generation systems and new power systems such as wind-solar-biomass multi-energy complementarity. This technology integrates several key steps, including biomass pretreatment, pyrolysis reaction control, and product separation, and demonstrates enormous application potential in energy production and environmental remediation. For example, in energy production, biomass co-pyrolysis can efficiently convert biomass into clean energy sources such as bio-oil and biogas; in environmental remediation, the biochar produced by co-pyrolysis can be used for soil improvement and pollutant adsorption, thereby significantly improving environmental quality. As biomass co-pyrolysis technology continues to mature, its key role in future sustainable energy and environmental management will become increasingly prominent, profoundly impacting the promotion of energy structure transformation and environmental protection.

[0003] Meanwhile, pyrolysis reaction control, as one of the core components of biomass co-pyrolysis technology, plays a crucial role, especially in fields such as bioenergy production, environmental remediation, and materials synthesis. Existing pyrolysis reaction control methods mainly rely on the precise regulation of reaction conditions, typically achieving efficient conversion of biomass into target products through the optimization of parameters such as temperature, pressure, and catalysts.

[0004] However, the amount of data available for predictive modeling of product yields during biomass co-pyrolysis is relatively limited, primarily due to the complexity of experimental conditions and cost constraints. Under specific reaction conditions, the efficiency and product selectivity of the pyrolysis reaction can fluctuate significantly, affecting its conversion effect and economic viability. This fluctuation mainly stems from the complexity of the reaction conditions. In high-temperature or hypoxic environments, the pyrolysis reaction of biomass may be inhibited, making complete conversion difficult and resulting in a decrease in product yield. Especially in areas with unsuitable reaction conditions, the energy in biomass is not fully utilized due to incomplete pyrolysis, further negatively impacting product quality and application scope. Furthermore, the rate and selectivity of the pyrolysis reaction may be limited in the absence of an effective catalyst. This characteristic not only reduces the purity and application value of the product but may also lead to the formation of byproducts, significantly increasing the difficulty and cost of subsequent separation and purification.

[0005] In existing technologies, Variational Autoencoders (VAEs) have demonstrated good performance in various data processing tasks, effectively handling high-dimensional data and complex data distributions. However, VAE performance can be significantly limited when faced with decreased prediction accuracy due to changes in data characteristics, especially with limited data volume. In contrast, Generative Adversarial Networks (GANs) can continuously update and optimize their models by generating new data, thus better adapting to changes in data characteristics. Through adversarial training between the generator and discriminator, GANs can generate new data similar to the real data distribution, which helps address the problem of data scarcity. However, GAN training can be lengthy, and it may struggle to quickly reach optimal performance under complex data distributions, limiting its application in scenarios requiring rapid response. Summary of the Invention

[0006] This invention addresses the challenges of data scarcity and abundant feedstock characteristics in biomass co-pyrolysis bio-oil yield prediction. It proposes a deep learning-based method for predicting renewable energy products, improving the model's adaptability and prediction accuracy under these conditions, thus enabling it to play a greater role in the biomass energy field. Specifically, this method combines the advantages of VAE and GAN. VAE rapidly extracts key features from biomass co-pyrolysis data, providing reliable prior information for GAN, significantly reducing training time and computational resource consumption. The GAN algorithm is used to generate new data samples, enabling the model to better understand and simulate the complex reaction process of biomass co-pyrolysis, further guiding the model to learn and predict in information-rich regions. The improved algorithm not only enhances the accuracy and reliability of the model in predicting biomass co-pyrolysis bio-oil yield but also strengthens the overall robustness and practicality of the system.

[0007] Implementation Step 1: Obtain the real dataset of biomass co-pyrolysis bio-oil yield, clean it to remove noise, missing values ​​and outliers, and normalize it; divide the preprocessed dataset into training dataset and test dataset.

[0008] Step 1.1: Normalize the preprocessed dataset, scaling the data to the [0, 1] interval to eliminate size differences between different features. The expression is:

[0009] ,

[0010] in, for Normalized value, For parameters The value, The minimum value of the parameter in the dataset This represents the maximum value of the parameter in the dataset.

[0011] Step 1.2: Use Pearson correlation analysis to evaluate the linear correlation between the features and the target variable on the normalized dataset, and screen out the parameters that are highly correlated with the target variable.

[0012] Implementation Step 2: Construct a VAE-GAN model to enhance the biomass feature dataset.

[0013] Step 2.1: VAE Feature Extraction. The VAE is trained using feature data from the training dataset to learn the latent features in the original data, and the model parameters are optimized by reconstructing the input data. The feature data includes experimental conditions such as pyrolysis temperature, raw material ratio, industrial analysis, and elemental analysis.

[0014] Step 2.1.1: The VAE network structure includes an encoder and a decoder. The encoder takes the input data... Mapping to the latent space, output distribution parameters:

[0015] ,

[0016] in, The posterior distribution of the encoder output. As latent variables, and These are the mean and variance of the output distribution parameters, respectively.

[0017] Decoder from latent variables Reconstructing input data The goal is to minimize the reconstruction error:

[0018] ,

[0019] in, This represents the likelihood distribution defined by the decoder. This represents the decoder network function in VAE.

[0020] Step 2.1.2: The loss function includes reconstruction loss and KL divergence. Reconstruction loss ensures that the decoder can reconstruct the input data, and KL divergence forces the latent space. Approximating the standard normal distribution:

[0021] ,

[0022] in, Indicates the encoder distribution Expectations Let KL divergence be the KL divergence.

[0023] After training, the raw data is input into the trained VAE encoder to extract features. .

[0024] Step 2.2: GAN Generates Feature Data. The generator of the GAN is trained using variables sampled from a standard normal distribution in a VAE; the data generated by the generator and... Train the discriminator. The discriminator distinguishes between these feature data and the real data, generating feature data that more closely approximates the distribution of the real data.

[0025] generator Features extracted from VAE Mapping to data space to generate fake data Discriminator Judge input data Is the data accurate?

[0026] The loss function formula for the discriminator is:

[0027] ,

[0028] in, It is the distribution of real data. Input the generator with a standard normal distribution of the VAE latent variables. It is the expectation of the actual data distribution. It is the expectation of the potential noise distribution.

[0029] The goal of the discriminator is to maximize the loss function:

[0030] ,

[0031] The generator loss function formula is:

[0032] ,

[0033] The generator's objective is to minimize the loss function:

[0034] ,

[0035] The joint optimization objective of GANs is to find a generator. and discriminator This makes the data distribution generated by the generator approximate the real data distribution.

[0036] ,

[0037] in, and These represent the input data. and The probability of being identified as real data.

[0038] Step 3: Construct a Support Vector Regression (SVR) model for predicting bio-oil yield and use Particle Swarm Optimization (PSO) to optimize the parameters to generate bio-oil yield data.

[0039] Step 3.1: Use the extracted feature data and its corresponding bio-oil yield data to train the SVR model to predict bio-oil yield. The goal of SVR is to find a hyperplane such that the error of all samples is constrained to a given value. Within this context, the SVR optimization problem can be represented as:

[0040] ,

[0041] Constraints:

[0042] ,

[0043] ,

[0044] ,

[0045] in, It is a weight vector. It is a bias term. Input data Feature vectors mapped to high-dimensional space and It is a slack variable. It is the penalty coefficient.

[0046] Using the Lagrange multiplier method to construct the Lagrange function to solve optimization problems:

[0047] ,

[0048] in, and It is a Lagrange multiplier.

[0049] right Taking the partial derivative and setting it to 0 yields the dual problem:

[0050] ,

[0051] ,

[0052] ,

[0053] in, It's a kernel function. It is a radial basis function. yes and The Euclidean distance between them is used to calculate the input data. and Inner product in higher-dimensional space.

[0054] Step 3.2: Use PSO to adjust the SVR model parameters - penalty coefficients and radial basis functions Optimize.

[0055] The objective of PSO is to find the optimal penalty coefficients and radial basis functions that minimize the mean squared error (MSE) of the SVR model. The objective function of PSO is defined as:

[0056] ,

[0057] in, It is the mean squared error of the SVR model on the training data.

[0058] First, randomly initialize the position and velocity of the particles.

[0059] ,

[0060] ,

[0061] Update the particle's velocity and position based on its individual optimal position and global optimal position:

[0062] ,

[0063] ,

[0064] in, It is a particle In time speed, It is a particle In time Location, It is a particle In time The optimal position of an individual It is the globally optimal position. It is the inertial weight, which controls the inertia of particle movement. and It is a learning factor that controls the movement of particles toward their individual optimal and global optimal positions. and It is a random number, and the range is... .

[0065] In each iteration, the SVR model is retrained based on the updated particle positions, and its MSE on the training data is calculated. By comparing the MSE values, the individual optimal position and the global optimal position are updated. Ultimately, the PSO algorithm converges to a set of optimal hyperparameter combinations, maximizing the predictive performance of the SVR model. The optimized SVR model is then used to input the feature data generated by the GAN, thereby generating the corresponding bio-oil yield. The generated bio-oil yield, together with the generated feature data, constitutes the synthetic data.

[0066] Step 4: Construct and train the final SVR model, and evaluate its performance on the test set. Mix synthetic data with real training data processed by VAE to form an enhanced training dataset, and train the final SVR model with optimized parameters. Then, extract features from the test data using the trained VAE encoder and input them into the final SVR model to obtain predicted bio-oil yields. Compare these predicted values ​​with the known real bio-oil yields in the test set, and finally evaluate the model's predictive performance using four metrics: MAE, MSE, RMSE, and R².

[0067] MAE indicator assessment:

[0068] ,

[0069] MSE indicator assessment:

[0070] ,

[0071] RMSE indicator assessment:

[0072] ,

[0073] R 2 Indicator Evaluation:

[0074] .

[0075] This invention presents a deep learning-based method for predicting renewable energy products, addressing the issue of limited datasets for predicting biomass co-pyrolysis products. Compared to existing technologies, the technical solution of this invention offers the following advantages.

[0076] 1. This invention combines the advantages of VAE and GAN, utilizing VAE to rapidly extract key features from biomass co-pyrolysis data, providing reliable prior information for GAN, thereby significantly reducing training time and computational resource consumption. In biomass co-pyrolysis experiments, real samples suffer from high experimental costs and large batch variations in raw materials. This method solves the model overfitting problem caused by the scarcity of co-pyrolysis experimental data. This approach not only improves the accuracy and reliability of the model in predicting bio-oil yield from biomass co-pyrolysis but also enhances the overall robustness and practicality of the system.

[0077] 2. This invention adaptively optimizes the hyperparameters of SVR using PSO. Targeting the data characteristics of "high-dimensional, nonlinear, and small sample size" in the co-pyrolysis of biomass and plastics, it effectively solves the nonlinear coupling problem in scenarios with a high plastic content, ensures prediction consistency across different raw material batches, reduces prediction errors, decreases computational resource consumption, and enhances model robustness. Attached Figure Description

[0078] Figure 1 This is an overall flowchart of an embodiment of the present invention. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0080] Table 1 Characteristic variables of biomass co-pyrolysis

[0081]

[0082] Appendix Figure 1 The above is an overall flowchart of an embodiment of the present invention. This embodiment provides a method for predicting renewable energy products based on deep learning. The specific process includes: acquiring a dataset and preprocessing it, enhancing bio-oil yield data based on VAE-GAN, constructing a bio-oil yield prediction SVR model, optimizing parameters using the PSO algorithm, and evaluating bio-oil yield prediction.

[0083] Implementation Step 1: Obtain the real dataset of biomass co-pyrolysis bio-oil yield, clean it to remove noise, missing values ​​and outliers, and normalize it; divide the preprocessed dataset into training dataset and test dataset.

[0084] Step 1.1: Normalize the preprocessed dataset, scaling the data to the [0, 1] interval to eliminate size differences between different features. The expression is:

[0085] ,

[0086] in, for Normalized value, For parameters The value, The minimum value of the parameter in the dataset This represents the maximum value of the parameter in the dataset.

[0087] Step 1.2: Use Pearson correlation analysis to evaluate the linear correlation between the features and the target variable on the normalized dataset, and screen out the parameters that are highly correlated with the target variable.

[0088] Implementation Step 2: Construct a VAE-GAN model to enhance the biomass feature dataset.

[0089] Step 2.1: VAE Feature Extraction. The VAE is trained using feature data from the training dataset to learn the latent features in the original data, and the model parameters are optimized by reconstructing the input data. The feature data includes experimental conditions such as pyrolysis temperature and raw material ratio, as shown in Table 1.

[0090] Step 2.1.1: The VAE network structure includes an encoder and a decoder. The encoder takes the input data... Mapping to the latent space, output distribution parameters:

[0091] ,

[0092] in, As latent variables, and These are the mean and variance of the output distribution parameters, respectively.

[0093] Decoder from latent variables Reconstructing input data The goal is to minimize the reconstruction error:

[0094] ,

[0095] in, This represents the decoder network function in VAE.

[0096] Step 2.1.2: The loss function includes reconstruction loss and KL divergence. Reconstruction loss ensures that the decoder can reconstruct the input data, and KL divergence forces the latent space. Approximating the standard normal distribution:

[0097] ,

[0098] in, Indicates the encoder distribution Expectations Let KL divergence be the KL divergence.

[0099] After training, the raw data is input into the trained VAE encoder to extract features. .

[0100] Step 2.2: GAN Generates Feature Data. The generator of the GAN is trained using variables sampled from a standard normal distribution in a VAE; the data generated by the generator and... Train the discriminator. The discriminator distinguishes between these feature data and the real data, generating feature data that more closely approximates the distribution of the real data.

[0101] generator Features extracted from VAE Mapping to data space to generate fake data Discriminator Judge input data Is the data accurate?

[0102] The loss function formula for the discriminator is:

[0103] ,

[0104] in, It is the distribution of real data. Input the generator with a standard normal distribution of the VAE latent variables. It is the expectation of the actual data distribution. It is the expectation of the potential noise distribution.

[0105] The goal of the discriminator is to maximize the loss function:

[0106] ,

[0107] The generator loss function formula is:

[0108] ,

[0109] The generator's objective is to minimize the loss function:

[0110] ,

[0111] The joint optimization objective of GANs is to find a generator. and discriminator This makes the data distribution generated by the generator approximate the real data distribution.

[0112] ,

[0113] in, and These represent the input data. and The probability of being identified as real data.

[0114] Step 3: Construct a Support Vector Regression (SVR) model for predicting bio-oil yield and use Particle Swarm Optimization (PSO) to optimize the parameters to generate bio-oil yield data.

[0115] Step 3.1: Use the extracted feature data and its corresponding bio-oil yield data to train the SVR model to predict bio-oil yield. The goal of SVR is to find a hyperplane such that the error of all samples is constrained to a given value. Within this context, the SVR optimization problem can be represented as:

[0116] ,

[0117] Constraints:

[0118] ,

[0119] ,

[0120] ,

[0121] in, It is a weight vector. It is a bias term. Input data Feature vectors mapped to high-dimensional space and It is a slack variable. It is the penalty coefficient.

[0122] Using the Lagrange multiplier method to construct the Lagrange function to solve optimization problems:

[0123] ,

[0124] in, and It is a Lagrange multiplier.

[0125] right Taking the partial derivative and setting it to 0 yields the dual problem:

[0126] ,

[0127] ,

[0128] ,

[0129] in, It's a kernel function. It is a radial basis function. yes and The Euclidean distance between them is used to calculate the input data. and Inner product in higher-dimensional space.

[0130] Step 3.2: Use PSO to adjust the SVR model parameters - penalty coefficients and radial basis functions Optimize.

[0131] The objective of PSO is to find the optimal penalty coefficients and radial basis functions that minimize the mean squared error (MSE) of the SVR model. The objective function of PSO is defined as:

[0132] ,

[0133] in, It is the mean squared error of the SVR model on the training data.

[0134] First, randomly initialize the position and velocity of the particles.

[0135] ,

[0136] ,

[0137] Update the particle's velocity and position based on its individual optimal position and global optimal position:

[0138] ,

[0139] ,

[0140] in, It is a particle In time speed, It is a particle In time Location, It is a particle In time The optimal position of an individual It is the globally optimal position. It is the inertial weight, which controls the inertia of particle movement. and It is a learning factor that controls the movement of particles toward their individual optimal and global optimal positions. and It is a random number, and the range is... .

[0141] In each iteration, the SVR model is retrained based on the updated particle positions, and its MSE on the training data is calculated. By comparing the MSE values, the individual optimal position and the global optimal position are updated. Ultimately, the PSO algorithm converges to a set of optimal hyperparameter combinations, maximizing the predictive performance of the SVR model. The optimized SVR model is then used to input the feature data generated by the GAN, thereby generating the corresponding bio-oil yield. These generated bio-oil yields, together with the generated feature data, constitute the synthetic data.

[0142] Step 4: Construct and train the final SVR model, and evaluate its performance on the test set. Mix synthetic data with real training data processed by VAE to form an enhanced training dataset, and train the final SVR model with optimized parameters. Then, extract features from the test data using the trained VAE encoder and input them into the final SVR model to obtain predicted bio-oil yields. Compare these predicted values ​​with the known real bio-oil yields in the test set, and finally evaluate the model's predictive performance using four metrics: MAE, MSE, RMSE, and R².

[0143] MAE indicator assessment:

[0144] ,

[0145] MSE indicator assessment:

[0146] ,

[0147] RMSE indicator assessment:

[0148] ,

[0149] R 2 Indicator Evaluation:

[0150] .

Claims

1. A deep learning-based method for predicting renewable energy products, characterized in that, Includes the following steps: Step 1: Obtain the real dataset of biomass co-pyrolysis bio-oil yield, clean it to remove noise, missing values ​​and outliers, and perform normalization. The preprocessed dataset is divided into a training dataset and a test dataset; Step 2: Construct a VAE-GAN model to enhance the biomass feature dataset; Step 3: Construct a bio-oil yield prediction SVR model and use the PSO algorithm to optimize the parameters to generate bio-oil yield data; Step 4: Build and train the final SVR model, and evaluate the performance of the final model on the test set.

2. The method for predicting renewable energy products based on deep learning according to claim 1, characterized in that, Step two involves constructing a VAE-GAN model to augment the biomass feature dataset, which is implemented according to the following steps: Step 2.1: VAE feature extraction; Use the feature data in the training dataset to train the VAE to learn the latent features in the original data, and optimize the model parameters by reconstructing the input data. The feature data are experimental conditions such as pyrolysis temperature and raw material ratio. Step 2.1.1: The VAE network structure includes an encoder and a decoder. The encoder takes the input data... Mapping to the latent space, output distribution parameters: ， in, As latent variables, and These are the mean and variance of the output distribution parameters, respectively. Decoder from latent variables Reconstructing input data The goal is to minimize the reconstruction error: ； Step 2.1.2: The loss function includes reconstruction loss and KL divergence. Reconstruction loss ensures that the decoder can reconstruct the input data, and KL divergence forces the latent space. Approximating the standard normal distribution: ， in, Indicates the encoder distribution Expectations Let KL divergence be the KL divergence. After training, the raw data is input into the trained VAE encoder to extract features. ; Step 2.2: GAN Generates Feature Data. The generator of the GAN is trained using variables sampled from a standard normal distribution in a VAE; the data generated by the generator and... Train the discriminator; the discriminator distinguishes between these feature data and real data, generating feature data that more closely approximates the distribution of real data; generator Features extracted from VAE Mapping to data space to generate fake data Discriminator Judge input data Is the data accurate? The loss function formula for the discriminator is: ， in, It is the distribution of real data. Input the generator with a standard normal distribution of the VAE latent variables. It is the expectation of the actual data distribution. It is the expectation of the potential noise distribution. The goal of the discriminator is to maximize the loss function: ， The generator loss function formula is: ， The generator's objective is to minimize the loss function: ， The joint optimization objective of GANs is to find a generator. and discriminator This makes the data distribution generated by the generator approximate the real data distribution. ， in, and These represent the input data. and The probability of being identified as real data.

3. The method for predicting renewable energy products based on deep learning according to claim 1, characterized in that, Step three involves constructing a bio-oil yield prediction SVR model and using the PSO algorithm for parameter optimization to generate bio-oil yield data. This is implemented according to the following steps: Step 3.1: Use the extracted feature data and its corresponding bio-oil yield data to train the SVR model to predict bio-oil yield; the goal of SVR is to find a hyperplane such that the error of all samples is constrained to a given value. Within this context, the SVR optimization problem can be represented as: ， The constraints are: ， ， ， in, It is a weight vector. It is a bias term. Input data Feature vectors mapped to high-dimensional space and It is a slack variable. It is the penalty coefficient; Using the Lagrange multiplier method to construct the Lagrange function to solve optimization problems: ， in, and They are Lagrange multipliers; right Taking the partial derivative and setting it to 0 yields the dual problem: ， ， ， in, It's a kernel function. It is a radial basis function. yes and The Euclidean distance between them is used to calculate the input data. and Inner product in higher-dimensional space; Step 3.2: Use PSO to adjust the SVR model parameters—penalty coefficients. and radial basis functions Optimize; The objective of PSO is to find the optimal penalty coefficients and radial basis functions that minimize the mean squared error (MSE) of the SVR model; the optimization objective function of PSO is defined as: ， in, It is the mean squared error of the SVR model on the training data; First, randomly initialize the position and velocity of the particles. ， ， Update the particle's velocity and position based on its individual optimal position and global optimal position: ， ， in, It is a particle In time speed, It is a particle In time Location, It is a particle In time The optimal position of an individual It is the globally optimal position. It is the inertial weight, which controls the inertia of particle movement. and It is a learning factor that controls the movement of particles toward their individual optimal and global optimal positions. and It is a random number, and the range is... ; In each iteration, the SVR model is retrained based on the updated particle positions, and its MSE on the training data is calculated. By comparing the MSE values, the individual optimal position and the global optimal position are updated. Finally, the PSO algorithm converges to a set of optimal hyperparameter combinations, enabling the SVR model to achieve the best prediction performance. The optimized SVR model is then used to input the feature data generated by GAN into the SVR model, thereby generating the corresponding bio-oil yield. These generated bio-oil yields, together with the generated feature data, constitute the synthetic data. MSE indicator assessment: 。

Images