Deep neural network and genetic algorithm synergistic optimization of new energy generation power combination prediction technology research

With the global concern for environmental pollution, new energy as a kind of environmentally friendly energy has received more and more attention. Power generation rate is an important index for the development of new energy field, and wind power and photovoltaic power generation are the two most common and representative forms of new energy, and the prediction of its energy generation rate becomes a key issue to ensure the stable operation of power grid and energy management [1–4]. Accurate prediction of the power output of new energy generation is crucial for achieving reliable dispatch and optimizing energy utilization. However, due to the uncertainties of wind, sunshine, complex data characteristics, seasonal changes, etc., it is difficult to accurately predict the energy generation rate, which has become a major problem in the development of new energy field [5–8].

For the power prediction of new energy generation, if the traditional statistical methods are used to analyze the numerical statistics of weather forecasts and historical data, the analysis results are inaccurate. According to the physical mapping relationship between weather and solar radiation intensity, it is necessary to transform the photovoltaic conversion model to predict the power [9–12], which is difficult to meet the requirements of the actual operational needs. In the past, it is mainly based on time, although the model is simple, but more and more new energy generation parameters are increasingly difficult to determine. And the deep neural network and genetic algorithm co-optimization can solve these problems [13–15].

A review of wind power forecasting, forecasting methods and progress is developed by Foley, A.M. et al. Statistical and machine learning methods are systematically introduced, the performance of the methods used in different prediction times is examined, and research activities, challenges and future developments are discussed [16]. Yu,F. et al. explore the impact of accurate prediction of new energy power generation on the consumption of new energy sources and traditional fossil energy sources based on new energy consumption operation mechanisms. Taking wind farms as the research object, the factors affecting the accuracy of short-term wind power prediction are examined, and strategies to improve the prediction accuracy of new energy power generation rate are outlined [17]. Zheng,J. et al. introduced a hybrid framework for the prediction of power generation of a variety of renewable energy sources, which utilizes CNN to extract the local correlation between the energy sources, and the validity of the framework is verified by taking the example of a renewable energy system, which shows that the hybrid framework other models are more accurate [18]. Widodo,D.A. et al. used LSTM as a learning model in order to provide accurate future predictions of electricity usage and renewable energy plants. And used confusion matrix accuracy value and RMSE error value as a prediction test [19]. Zhou,C.et al. developed a narrative on the factors affecting wind power with the aim of improving the quality of wind power. By studying the prediction methods for wind power prediction and optimizing the neural network using genetic algorithm to predict the wind power of a wind farm. It has certain reference significance for the power generation and grid connection of wind farms [20]. Chen,G. created a wind power prediction model combining CNN and GA based on the development of wind power generation and the current research status of wind power prediction technology. The experimental results emphasize that this prediction system has higher prediction accuracy and stability relative to other methods such as CNN [21].

In this paper, the training of the deep confidence network structure is effectively implemented by stacking the training of RBM as the pre-training part of the whole DBN and setting the BP network in the last layer for fine-tuning. After the training is completed, the K-means classifier is set to extract data features. The Adam algorithm is introduced to optimize the combination, increase the independence of parameters, and improve the training speed and stability of the model. Using a genetic algorithm, the initial weight matrix and threshold of the model are obtained to form the initial DBN network, and the GA-DBN model is established. The model achieves the prediction of new energy generation power by establishing a mapping relationship between input and output data. Taking the wind power farm in a province as the experimental object, the model predicts the whole new energy power generation process after the meteorological feature selection is completed, and analyzes the change rule, prediction accuracy and the actual power dispatching application.

2

Content and principles of research on power mix forecasting for new energy generation

2.1

New energy generation

2.1.1

Combined forecasting models

Combined prediction model is to combine two and more single prediction models in a specific way to form a comprehensive set of combined prediction methods [22]. There are many combined prediction methods, but each of them has its own advantages and disadvantages, and their prediction effects may be very different under different data conditions, but there is also a certain connection between them, and they can complement each other. In the combination of prediction methods, if the error of a single prediction model is very large, you can choose to eliminate it and use the most appropriate combination of several single models.

A large error in a single forecasting model in a combined forecasting model will not affect the whole to a high degree. The decision maker uses a single predictive model to make the prediction. If the selected model does not match the data at all, then the error in the predicted value will be large and may have a significant impact on the decision. If a combination of forecasting methods is used, the error value will not be too large. Therefore, the combination of prediction methods can further improve the accuracy and reliability of data prediction on the basis of a single prediction method. In recent years, academics have paid increasing attention to the combination prediction method and have achieved certain results.

2.1.2

Rationale for model combination practice

Combination of forecasting methods first need to get a lot of “single prediction model”, from which the judgment and screening, remove some obviously can not achieve a better effect of the prediction effect of the prediction model, and then screened out a number of “single prediction model” for comprehensive analysis Subsequently, a number of “single prediction models” screened out are analyzed comprehensively, and the results of different prediction models are combined reasonably and scientifically, so as to get the unique combination model with the best prediction effect at the same time. Finally, the weight coefficients of each model are calculated using the specific combination prediction method, and the final power prediction result is calculated.

2.2

New Energy Power Forecasting Implementation Program

This project has a good foundation for implementation. The power company has carried out a lot of research in the areas of refined meteorological forecasting technology, new energy power prediction and evaluation, and assessment of the two rules, and has accumulated practical experience in theoretical research and application engineering. This project fully utilizes the existing research results and is implemented in the following stages.

2.2.1

Analysis of New Energy Forecasting Algorithms and Applications

Fully investigate the current status of wind power and photovoltaic power prediction methods and applications at home and abroad, and organize the main prediction algorithms and classifications of mainstream wind power and photovoltaic at home and abroad, as well as the application scenarios.

2.2.2

Research on new energy power mix prediction methods

We investigate the current situation of combination prediction methods at home and abroad, study the current mainstream new energy power combination prediction methods, including new energy power combination prediction methods based on ensemble empirical decomposition and deep learning methods, short-term new energy power combination prediction methods based on empirical modal decomposition and support vector machines, and combination prediction methods based on weighted combinations, and compare and analyze the advantages and disadvantages of the various methods, and propose a new energy power combination prediction method with more accurate and more effective calculation. A new method for predicting energy power combinations with more accurate and effective calculations is proposed to achieve dynamic calculation of combination weights in different time periods.

2.2.3

Development of Combined Prediction Algorithms

Develop a new algorithm for predicting energy power combinations using multi-algorithm access, and optimize the combination prediction model algorithm by calculating and assessing case data quantitatively.

3

Deep Confidence Network Structure and Optimization

3.1

Deep Confidence Network Structure

A deep confidence network consists of several layers of Restricted Boltzmann Machines (RBMs) stacked on top of each other, and the model of an RBM is shown in Fig. 1 [23]. Its nodes include both visible and implied nodes, with bidirectional symmetric connections between the nodes in the visible and input layers, and no connections between the nodes in the same layer. The DBN consists of a visible layer, an implied layer, and an output layer, and the visible layer of the first RBM is the visible layer of the DBN, with the output of the previous RBM serving as the input to the next RBM. The traditional RBM can only handle binary data, while the neurons of Continuous Restricted Boltzmann Machine (CRBM) have the advantage of being able to continuously change state values and can handle real-valued data, and the DBN in this paper adopts CRBM.

Where v represents the visible layer unit and h represents the hidden layer unit, both of which are Bernoulli distributed. Therefore these nodes have two states: state 1 means that the node is activated, and state 0 means that the node is not activated, and the model will choose the node that is in the activated state to be used. The process of calculating the energy function of the RBM in the neuron’s corresponding state T = (V, H) is shown in Equation (1): 1 $E (V, H) = - \sum_{i = 1}^{n} a_{i} v_{i} - \sum_{j = 1}^{m} b_{j} h_{j} - \sum_{i = 1}^{n} \sum_{j = 1}^{m} v_{i} w_{i j} h_{j}$

Where a_i is the i th visible layer unit bias Kan, b_j is the j th hidden layer unit bias threshold, w_ij is the weight between the i th visible layer unit and the j th hidden layer unit, and v_i and h_j denote the activation state (0 or 1) of the visible and hidden layer neurons, respectively.

The energy function represents the stable state of each layer of RBM, in short, it is the energy model to measure the energy of the system in a certain state, intuitively speaking, the energy corresponding to each state describes the probability that the system is in this state, by the common sense of the natural sciences, the smaller the energy of a substance, the more stable the substance is. The point of training the RBM is to adjust the internal parameters of the model so as to fit a known training sample, so that the energy function gradually decreases and stabilizes, so that the probability distribution of neuron nodes of the RBM under the parameter conditions is as similar as possible to that of the training sample.

In general, it is not possible to fit the distribution of a data by the energy function alone, so for training samples with arbitrary unknown distributions, it is still necessary to fit them to the form of probability distributions by the energy function, which is calculated as shown in (2): 2 $P = \frac{e^{- E (1^{*}, H)}}{Z (θ)}$ where $Z (θ) = \sum_{v . h} e^{- E : (V . H)}$ is the normalization factor.

This probability distribution function is essentially the loss function of the network, which can also be thought of as the result of normalizing the energy function after exponentiating it. This realizes that a state of the training sample after training through the network corresponds to an energy, and the corresponding probability distribution corresponds to the probability distribution of the energy. The larger the probability value is the smaller the energy function is, i.e., the network fits better.

Due to the special inter-layer interconnection of RBM, the intra-layer is constructed without connection. For a given visible unit state, the activation probability of the hidden layer unit connected to it is shown in Equation (3): 3 $P (h_{j} = 1 | v) = δ (b_{j} + \sum_{i} v_{i} w_{i j})$

Under the corresponding conditions, the activation probability of a visible layer cell is shown in Equation (4) for a given hidden layer cell: 4 $P (v_{i} = 1 | h) = δ (a_{i} + \sum_{i} w_{i} h_{i})$ where $δ (x) = \frac{1}{1 + e^{- x}}$ represents the sigmod function, which is also a commonly used activation function for neural networks.

By setting a random number (between 0 and 1) to each node and then comparing the size of the random number with the activation probability, the unit is activated to participate in the model training when the random number is greater than the activation function, otherwise it is not activated.

After understanding the working principle of RBM above, in the actual data analysis, DBN is to achieve the purpose of learning data features by stacking RBM and training layer by layer, as can be seen from the structure of the RBM, each layer of the RBM has five parameters: v,h,b,c,w, where b,c,w corresponds to the weight and bias value of each layer of the network, which is learned by the network after inputting the training samples. v then denotes the input value of each layer of the network, Table v in the hidden layer in DBN indicates the activation state of the neurons, and v₁ in the input layer indicates the input training samples. Accordingly, h represents the output value of each layer of the network, h of the hidden layer in DBN indicates the activation state of the neuron in the next input layer, and h_m of the output layer (assuming that the number of layers of the deep confidence network is n) represents the output value, which can be the regression prediction value, or the fitted value according to the features of the input data compressed and then decompressed, or the feature value of the input data extracted from the input data after going through the neural network. This value can be a regression prediction value, or a fitted value based on the input data features compressed and decompressed, or a feature value extracted from the input data after the neural network. The details depend on the specific application of the network.

In practice, Gibbs sampling is generally used to estimate P, and the maximum likelihood function method is used to make the likelihood function constantly maximized and the energy function minimized. However, this method is cumbersome and inefficient.

For a random variable x of the input network, the two-layer RBM training process, for example, is implemented in the following steps: 1)

After initializing the network parameters, the x values (0 < x < 1) normalized to the training samples are assigned to the network input layer v₁, and the activation probability P(h₁ | v₁) of each neuron in the hidden layer is calculated through equation (3).

2)

Select the activated neurons in the hidden layer, reconstruct the next RBM’s questioning layer using its bias value b₁, and compute the activation probability P(v₂ | h₁) of each neuron in the hidden layer of the next RBM through equation (4).

3)

Similarly, continue to select the bias value b₂ of the activated neuron and calculate the activation probability of the hidden layer to obtain the probability distribution P(h₂ | v₂) of the hidden layer at this point.

4)

From this the network weights can be updated and the update formula is shown below: 5 $w = w + l (P (h_{1} | v_{1}) b_{1} - P (h_{2} | v_{2}) b_{2})$ 6 $a = a + l (v_{1} - b_{1})$ 7 $b = b + l (b_{1} - b_{2})$

According to the above, the training of RBM is an unsupervised learning process without labeling information, and the training of RBM by stacking RBM serves as the pre-training part of the whole DBN, and in general, a BP network is set up in the last layer of the DBN (DBN-BP), and the output feature vectors of the previous RBM layer are used as its input feature vectors to fine-tune the network and supervise the training of the network through error backpropagation network. In practice, depending on the specific application area, it can be replaced with any classifier model, not necessarily a BP network. In this paper, since DBN is needed to extract data features to complete clustering, a K-means classifier is set after RBM, and accordingly, when power prediction is performed, a BP network is set at the last layer for regression prediction.

3.2

Optimization of Deep Confidence Networks

All of the above optimization algorithms have their own advantages and can be combined for different parameters. Adam’s algorithm is a combination optimization algorithm proposed based on the ideas mentioned above [24]. For each parameter, it not only has its own learning rate, but also its own momentum, in such a training process, the update of each parameter is more independent, which improves the training speed and stability of the model. Its overall realization form is shown in the following equation: 8 $V (t) = m V (t - 1) + (1 - m) Δ j (g)$ 9 $d_{i} = p d_{i - 1} + (1 - p) Δ j (^{g) 2}$ \begin{eqnarray}{d}_{i}=p{d}_{i-1}+(1-p)\Delta j{(}^{g}\end{eqnarray} 10 $V (t) = \frac{V (t)}{1 - m}$ $$V(t) = {{V\left( t \right)} \over {1 - m}}$$ 11 ${\hat{d}}_{i} = \frac{d_{i}}{1 - p}$ 12 $g_{i} = g_{i - 1} - \frac{l}{\sqrt{{\hat{d}}_{i} + ε}} \hat{V} (t)$

4

New energy power prediction models

4.1

GA-DBN prediction model design

4.1.1

Application of genetic algorithms

GA is an optimization algorithm based on the principle of natural evolution. In the population consisting of the parameters to be optimized, individuals are screened according to the selected fitness function, and operations such as cross-mutation in genetics are carried out. The probability that an individual produces offspring is directly proportional to its fitness, and thus in the process of continuous evolutionary screening, the offspring will have an increasing fitness function value based on the inheritance of the previous generation. When the end condition is satisfied, the individual with the highest value of individual fitness is selected. The process of training the initial value of a DBN neural network using GA is shown in Fig. 2, and the specific steps are as follows [25]. 1)

Population initialization. Individuals are encoded using real number encoding, which can avoid complex encoding and decoding process, and each individual is a real number string, which consists of the connection weights of input and implicit layers, implicit and implicit layers, implicit and output layers, and the thresholds of implicit and output layers.

2)

The fitness function is chosen as: 13 $F_{i} = 1 / E_{i}$ where E_i is the error between the predicted output and the desired output after learning by the i th individual.

3)

Gene selection. The adaptation values obtained for each individual were arranged in descending order and manipulated using roulette wheel selection to obtain the probability of their occurrence in the offspring, the higher the adaptation the greater the probability of inheritance to the offspring, and the probability of individual i being selected was: 14 $P_{i} = F_{i} / \sum_{i = 1}^{c} F_{i}$ where is the total number of individuals.

4)

Gene crossover. The real crossover method is used to generate two new individuals by exchanging some of the genes between two paired individuals with probability p_c. Individuals m and n are crossed at position j as follows, i.e.: 15 ${\begin{array}{l} g_{m j}^{'} = g_{m j} r + g_{n j} (1 - r) \\ g_{n j}^{'} = g_{n j} r + g_{m j} (1 - r) \end{array}$ where g_mj, g_nj is the gene r ∈ [0,1] at position j for the m th and n th individuals, respectively, before crossover.

5)

Genetic mutation. In order to increase the diversity of individuals in a population, mutation is performed with a low probability p_m. The j th gene of the m th individual was randomly selected for mutation and the operation was performed as follows, i.e.: 16 $g_{m j}^{'} = {\begin{matrix} g_{m j} r + (g_{m j} - g_{\max}) r_{1} (1 - s / s_{\max}) r_{2} \geq 0.5 \\ g_{m j} r + (g_{\min} - g_{m j}) r_{1} (1 - s / s_{\max}) r_{2} < 0.5 \end{matrix}$ where g_min, g_max is the minimum and maximum values at the j locus of the m th individual, respectively, s is the current number of iterations, s_max is the maximum number of evolutions, r₁ is a random number, and r₂ ∈ [0,1].

6)

Judge whether the termination condition is satisfied. After the mutation operation, if the set maximum number of iterations is reached or the calculated fitness function value reaches the set accuracy, the individual with the highest fitness among them is output, otherwise the next round of replication, crossover or mutation operation is carried out until the condition is satisfied.

4.1.2

GA-DBN modeling

The initial weight matrix and threshold values obtained from GA are utilized to form the initial DBN network. After that the training phase of DBN is carried out as follows [26]. 1)

Let the node state be denoted by {S_i}, and the result of connecting with node i (the state of point j be denoted by S_j, and the weight matrix be W. randomly select a training sample to input into the first CRBM (i.e., the visible layer of the DBN), and update the state of each node of the first implicit layer, S_j, i.e: 17 $S_{j} = φ_{j} (\sum_{i} ω_{i j} S_{i} + σ N_{j} (0, 1))$ where φ_j(x) is a s-type function and N_j(0,1) is the standard normal distribution.

2)

Based on the implied node state S_j obtained in step 1), update the visible node state according to Eq. (12), denoted as S_i, i.e: 18 $S_{i} = φ_{j} (\sum_{j} ω_{i j} S_{j} + σ N_{i} (0, 1))$

3)

Based on the visible node state S_i updated in step 2), update the implied node state again according to Eq. (18), denoted as S_j, i.e: 19 $S_{j} = φ_{j} (\sum_{i} ω_{i j} S_{i} + σ N_{j} (0, 1))$

4)

Use the implicit layer node state S_j of the first CRBM as the input to the second CRBM, and then update the state of each node of the second CRBM in accordance with steps 1) to 3). Follow this step layer by layer until the node state of the whole network is updated.

5)

Randomly select the next training sample and jump to step 1). When all the training samples are input, this round of training is finished, then the amount of change in each weight value is calculated according to Eq. (19) in order to update the weight matrix of each CRBM, i.e: 20 $Δ ω_{i j} = η_{ω} (< S_{i} S_{j} > - < S_{i} S_{j} >)$

6)

Go to step 2) for the next round of training until the set maximum number of trainings is reached or the amount of change in the weight matrix of the network is small enough, i.e., when ||DW(i)||< e, the DBN training is ended and the DBN network modeling is completed.

4.2

Predictive modeling

The prediction method in this paper is mainly to establish mapping relationship between input data (wind speed, temperature, etc.) and output data (wind power) through the prediction model to achieve the prediction purpose. one day from the test set was selected as the sample to be tested, and according to the results of the clustering model and the discriminant analysis to determine the data category to be tested belongs to the data category of the day to be tested, and then the data of the wind speed and temperature of the date of the category of data was normalized to the input to train the prediction model. the structure of the prediction model is shown in Fig. 3. The prediction model’s structure is depicted in Fig. 3.

The specific steps are as follows: 1)

Preparation of training samples: after data clustering analysis, the historical data is divided into several classes of cluster data, and its category is determined by the discriminant analysis of the day to be measured, and the wind power data of this category is used as the training samples.

2)

Model parameter adjustment: two-layer stacked RBM is used, and finally the structure of BP feedback network is set, the initial learning rate is 0.1, the number of iterations is 300, the input nodes are 2 (i.e., wind speed and temperature), and after optimization by genetic algorithm, the first implicit layer node is 44, the second implicit layer node is 72, and the output layer node is 1 (i.e., output wind power). The network weights are initialized to 0, and the network bias and weights are continuously updated after error back propagation until the error is less than a given value (0.001).

3)

Short-term wind power prediction: when the model is trained, input the wind power data of the day to be measured, and then back-normalize the output results to get the 24-hour prediction power of the day.

5

Experimental analysis of new energy generation power prediction

5.1

Data Acquisition

5.1.1

Historical meteorological data of wind farms

A province wind power farm as an experimental object, the total installed capacity of the farm is 273,960,000 kilowatts, of which, the installed capacity of wind power is 209,150,000 kilowatts, photovoltaic installed capacity is 9 kilowatts, the wind power project covers an area of about 756,156.454 square meters, photovoltaic project covers an area of about 709,165.651 square meters, and the historical meteorological data are shown in Table 1.

Table 1.

Historical meteorological data of landscape hair field

Meteorological data category	Numerical value
Annual average temperature/°C	10.165
Annual maximum temperature/°C	42.899
Annual minimum temperature/°C	-25.563
Average annual relative humidity (%)	42.955
Annual average wind speed(m/s)	2.948
Maximum annual wind speed(m/s)	15.866
Annual average pressure/hPa	942.856
Annual average precipitation/mm	37.511
Maximum annual precipitation/mm	26.685

The GA-DBN new energy power prediction model is used to collect the wind and photovoltaic generator operation data and environmental data of the wind farm shown in Table 1, and take the current data of the wind generator and the wind speed data as an example to analyze the data collection effect of the GA-DBN new energy power prediction model, and the results of the data collection are shown in Figure 4. According to Fig. 4, it can be seen that the GA-DBN new energy power prediction model can effectively collect the current data of wind turbine and the wind speed data of wind farm, the current of wind turbine fluctuates between -200~200A, and the wind speed is between 0.3~3.7m/s. The data collection results can be seen in Fig. 4. It shows that the GA-DBN new energy power prediction model has the effectiveness of collecting wind and photovoltaic generator operation data, as well as environmental data, which can provide effective data support for the subsequent wind power prediction.

5.1.2

Selection of meteorological features

Considering that each group of meteorological variables has a certain degree of randomness and is characterized by normal distribution, the Pearson correlation coefficient is chosen in this paper to describe the degree of correlation between a variety of meteorological characteristics and the actual power generation, and is visualized by the Pearson heat map. Its expression is: 21 $ρ_{X, Y} = \frac{cov (X, Y)}{σ_{X} σ_{Y}} = \frac{E [(X - μ_{X}) (Y - μ_{Y})]}{σ_{X} σ_{Y}}$

Where: the Pearson’s correlation coefficient between X and Y is denoted by ρ_X,Y, cov(X,Y) is the covariance of X and Y, σ_X and σ_Y are the standard deviation of X and Y, respectively, E is the expected value, and μ_X and μ_Y are the means of X and Y, respectively.

Equation (21) is commonly used in the field of new energy forecasting to calculate the Pearson’s correlation coefficient between the meteorological variables X and the power Y for a certain time span to characterize the correlation between the meteorological characteristics and the new energy power generation power in that period of time. 22 $γ_{i} = \frac{\sum_{j = 1}^{n} (X_{i j} - {\bar{X}}_{i}) (Y_{j} - \bar{Y})}{\sqrt{\sum_{j = 1}^{n} {(X_{i j} - {\bar{X}}_{i})}^{2} \sum_{j = 1}^{n} {(Y_{j} - \bar{Y})}^{2}}}$ γ_i is the correlation coefficient between i and actual power generation in a certain period of time, and its value ranges from (-1,1). γ_i is greater than 0 means positive correlation, less than 0 is negative correlation, the absolute value is closer to 1 means the stronger correlation. X_ij denotes the recorded value of the meteorological feature at the ith moment in the statistical period. j denotes the actual generation power corresponding to tha t moment. Y_j denotes the actual generation power corresponding to that moment. ${\bar{X}}_{i}$ and $\bar{Y}$ denote the mean value of the meteorological characteristic i and the mean value of the actual generation power during the time period, respectively. n Indicates the number of records in the statistical time period.

The original dataset was first cleaned to eliminate duplicate values, fill in the blank values and remove outliers, and then normalized. The Pearson correlation coefficient is chosen to describe the degree of correlation between meteorological features and the actual power of wind power, and the correlation heat map is plotted, and the results are shown in Figure 5.

From the thermodynamic diagram, it can be seen that the correlation between 10m, 30m, 50m and hub wind speed on the actual power of wind power is high, so it is selected as the input feature of the prediction model in this paper.

5.2

Analysis of new energy generation power forecast

5.2.1

Forecasting process

In order to reduce the impact of the random volatility of wind power generation and to improve the accuracy of the GA-DBN network’s prediction results of wind power generation, this paper adopts the EMD method to decompose each pair of signals after adding white noise to obtain the IMF component and the residual RE corresponding to each pair of signals, and performs the set-total averaging for each component, i.e.: 23 $I M F_{j} = \frac{1}{2 N} \sum_{i = 1}^{N} (I M F_{i j} + I M F_{- i j})$ 24 $R E_{j} = \frac{1}{2 N} \sum_{i = 1}^{N} (R E_{i j} + R E_{- i j})$ where IMF_ij is the j th IMF component of the decomposition obtained by adding positive white noise at the i th time, IMF_–ij is the j th IMF component of the decomposition obtained by adding negative white noise at the i th time, and IMF_–ij is the j th IMF component of the original signal. RE_ij is the jth residual term RE obtained from the decomposition of the ith addition of positive white noise, RE_–ij is the j th residual term RE obtained from the decomposition of the i th addition of negative white noise, and RE_j is the j th residual term RE of the original signal.

This test uses historical real data from January 1, 2021-December 31, 2022 as the training set and January 2023 data as the test set.

The PV output power is 0 at night, so it is necessary to exclude the data of the 0 part of the power at night, and according to the actual data, intercept the power data of 07:00-18:00 for a total of 11 h as the experimental value.

Firstly, the average power per 1 h is obtained from the average power per 15 min, and secondly a different positive and negative white noise is added to the initial hourly PV power data each time for a total of three times. Finally, empirical modal decomposition is performed for each of the six sets of data, and the partial decomposition results for one set of data are shown in Fig. 6.

As can be seen from Fig. 6, sub-IMF₁ and IMF₂ are random components, this part of the frequency is high, the value is more difficult to predict, the prediction accuracy is low but the amplitude change is small, the amplitude of the change is between 0~0.6 and -0.2~0.4, respectively. Therefore, it does not have much influence on the final results. IMF₃ ~ IMF₈ is the periodic component, which has low frequency, shows a certain periodic pattern of changes, has higher prediction accuracy and larger amplitude changes, and thus has a greater influence on the final results. RE is the residual term, and the upward trend from -0.1 to 0.1 indicates an overall increasing trend of PV power in this component of the results. For each component of the results, train its corresponding GA-DBN network.

5.2.2

Analysis of the pattern of change in new energy power generation

In order to ensure the universality and fairness of the results of the study, to expand the scope of analysis, from the collection of historical power data and meteorological data in the collection of sunny, cloudy, rainy day 3 types of weather randomly selected historical days, and each type of weather under the selection of 10 days of historical power data, calculate the same type of weather conditions in each moment of the average value of power generation for the depiction of the power generation of each type of weather change law curve. In view of the high reference value of power generation from 8:00 to 17:00 in a day, the horizontal coordinate time is chosen as the whole point of time in this period, and the power generation area of photovoltaic power plant under each type of weather is shown in Fig. 7.

Analyzing Fig. 7, it can be seen that the generation power of the historical day under sunny conditions changes synchronously with the change of solar radiation intensity, and the generation power is higher and relatively stable at 10:00-15:00, i.e., the generation power reaches the maximum of the day during the midday time period, and the generation power ranges from 300 to 375 kW. Rainy day power generation is affected by rainfall, rainfall time period and other factors to produce a large degree of fluctuation, while cloudy day power generation by cloud and ambient temperature and other factors, the overall show small fluctuation state, affected by the type of weather, the daily power generation has the corresponding rule of change.

5.2.3

Power Dispatch Applications

Due to the economic development of the experimental area is relatively backward and the number of population is not large, according to the official data statistics, the city in May 2021 daily electricity load demand fluctuation amplitude is not large, so the method of this paper is used to the power station in a month of power generation power prediction, and to take the corresponding power scheduling measures, power scheduling details are shown in Figure 8.

Analysis of Figure 8 shows that, after obtaining the prediction curve of power generation of PV power plant, the relevant departments discuss and make the corresponding power dispatch decision, which is just able to meet the daily power load demand of the city, based on the new energy power generation demand predicted by the GA-DBN model, the power value of power dispatch output is between 20~205kw, and the power that needs to be dispatched is mainly concentrated in the power of less than 100kw, and the demand of more than 200kw accounts for fewer. The power to be dispatched mostly focuses on power below 100 kW, and the demand above 200 kW accounts for less. The difference between the actual value of PV power and the predicted power is not large, and there is no excessive waste or shortage of power resources, which proves that the method in this paper can accurately predict the power generated by PV power plants, and thus contribute to the smooth operation of power scheduling.

5.2.4

Analysis of the accuracy of power generation forecasts

In order to compare the prediction effect of each model, this chapter uses three commonly used prediction accuracy evaluation indexes to assess the model prediction results. 1)

Mean square error, the expression of which is as follows: 25 $M S E = \frac{1}{N} \sum_{i = 1}^{N} {(P_{r, i} - P_{f, i})}^{2}$

Where, MSE is the mean square error, N is the prediction step size, P_f,i is the power prediction result at step i, and P_τ,i is the true value of power at step i. the smaller the MSE, the higher the accuracy of the algorithm’s prediction.

2)

Mean absolute error, which is expressed as follows: 26 $M A E = \frac{1}{N} \sum_{i = 1}^{N} | (P_{r, i} - P_{f, i}) |$

Where MAE is the average absolute error, N is the prediction step size, P_f,l is the power prediction result at step i, and P_f,l is the true value of power at step i. The smaller MAE is, the higher the accuracy of the algorithm’s prediction.

3)

Root-mean-square error, which is expressed as follows: 27 $R M S E = \sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(P_{r, i} - P_{f, i})}^{2}}$

Where RMSE is the root mean square error, N is the prediction step size, P_f,l is the power prediction result at step i, and P_r,l is the true value of power at step i. the smaller RMSE is, the more accurate the prediction of the algorithm is.

The models in this paper and the LSTM model with the same parameters of the corresponding module, the GRU model, the BiGRU model, the AM-BiGRU model, the TPA-BiGRU model, the CNN&TPA-BiGRU model using the Stacking method, and the DC-CNN&TPA-BiGRU model using simple linear regression (LR) are used to predict wind power in the next 15 min, and the results of each model are shown in Table 2. The wind power in the next 15 min is predicted by the CNN&TPA-BiGRU model using Stacking method and the DC-CNN&TPA-BiGRU model using simple linear regression (LR).

Table 2.

The prediction results of each model in the future 15min

Model	eMAE	eRMSE	eMSE	_R2
LSTM	0.0378	0.0559	0.0752	0.8269
GRU	0.0359	0.0533	0.0821	0.8345
BiGRU	0.0285	0.0485	0.0715	0.8963
AM-BiGRU	0.0245	0.0405	0.0592	0.9254
TPA-BiGRU	0.0286	0.0398	0.0654	0.9345
CNN&TPA-BiGRU(EL)	0.0208	0.0349	0.0741	0.9425
DC-CNN&TPA-BiGRU(LR)	0.0215	0.0352	0.0685	0.9454
GA-DBN	0.0205	0.0342	0.0545	0.9525

The GA-DBN model proposed in this paper, based on the above comparative models, has e_MAE reduced by 45.77%, 42.9%, 28.07%, 16.33%, 28.32%, 1.44%, and 4.65%, respectively, with a e_RMSE value of 0.0342 and a R² value of 0.9525, which are all ranked first.

Figure 9 shows the comparison of the model prediction curves, in which the values of 24-hour wind power predicted by four of the algorithms are clearly shown and compared with the actual value of wind power. From the figure, it can be clearly seen that the GA-DBN model proposed in this paper is more similar to the fluctuation of the curve of the real value of wind power compared with the other algorithms, which indicates that the method proposed in this paper is more suitable for wind power prediction with higher accuracy. Meanwhile, the overall trend is closer to the real curve, and the overall maximum error does not exceed ±0.2, which further indicates that the model proposed in this paper is effective for new energy generation power prediction.

6

Conclusion

In this paper, the content and principles of the prediction of new energy power generation combined model are analyzed, and the corresponding implementation scheme is proposed. On the basis of a deep confidence network, a genetic algorithm is added to establish a GA-DBN model. The prediction of new energy power generation is achieved by establishing a mapping relationship with input data of the prediction model. The EMD method is used to add white noise to decompose the wind power generation power, and the variation amplitudes of IMF₁ and IMF₂ are between 0~0.6 and -0.2~0.4, respectively. The residual term shows the trend from -0.1~0.1 upward indicates that the PV power in this part of the result shows an overall growth trend, according to the decomposition results, train the corresponding GA-DBN network. The new energy power generation power change rule is analyzed, under the condition of sunny days, the new energy power generation power in the time period of 10:00-15:00 time period is the most power generation power in a day, between 300~375kW, while cloudy days and rainy days as a whole present a small range of fluctuation state, the new energy power generation power change rule is affected by the type of weather, and the daily power generation has the corresponding change pattern. After obtaining the prediction curve of photovoltaic power station power generation, the relevant departments discuss and make the corresponding power scheduling decision, the power value of power scheduling output is between 20~205kw, and the power that needs to be dispatched is mainly concentrated in the following 100kw, which proves that the method of this paper can quasi-predict the power of photovoltaic power station power generation, and ensure the smooth operation of the power scheduling.

Language:: English

Publication timeframe:: 1 times per year
Journal Subjects:: Life Sciences, Life Sciences, other, Mathematics, Applied Mathematics, General Mathematics, Physics, Physics, other

Journal RSS Feed

Deep neural network and genetic algorithm synergistic optimization of new energy generation power combination prediction technology research

Zhongyuan Yan

Wulei Xue

Yi Zhang

Xinyi Du

Xiande Zheng

Published Online: Mar 19, 2025

Received: Nov 06, 2024

Accepted: Feb 21, 2025

DOI: https://doi.org/10.2478/amns-2025-0414

KeywordsAdam algorithm, GA-DBN model, Deep confidence network, Boltzmann machine, New energy generation power prediction

© 2025 Zhongyuan Yan et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Keywords
Adam algorithm, GA-DBN model, Deep confidence network, Boltzmann machine, New energy generation power prediction