A new energy output scene generation method and system based on a deep belief network

Abstract

The application provides a new energy output scene generation method and system based on a deep belief network, comprising: performing forward training on new energy historical output data in a constructed deep belief network, initializing a network learning rate, and generating new energy output training data; inputting the new energy output training data and feature data labels into a logistic regression layer of the deep belief network, correcting the feature data in the new energy output training data based on the network learning rate using a gradient descent method, and outputting new energy output training result data; and calculating a network error, when the network error is less than a network error threshold, outputting the new energy output training result data as a new energy output scene; the application deeply mines various characteristics of new energy historical output data, the obtained output scene is more scientific and reasonable, and is applied to generation of multiple new energy output scenes, thereby reducing difficulties in modeling faced in generation of new energy medium and long term output scenes.

Description

Technical Field

[0001] This invention belongs to the field of power technology, specifically relating to a method and system for generating new energy output scenarios based on deep belief networks. Background Technology

[0002] Stochastic optimization programming of renewable energy in power systems is typically conducted on medium- to long-term timescales, and long-term renewable energy output time series modeling is a crucial foundation for such programming. Renewable energy output exhibits stochastic fluctuations, which must be considered in long-term output time series modeling. Existing research often employs scenario generation methods to simulate the stochasticity of renewable energy output. This involves randomly sampling renewable energy output to obtain possible scenarios and using multiple deterministic scenarios to characterize the stochasticity. These methods primarily generate renewable energy scenarios on short-term timescales and are difficult to directly apply to generating long-term renewable energy output scenarios. Existing methods for modeling long-term renewable energy output sequences mainly include autoregressive moving average models, fluctuation-based modeling methods, and Markov-Monte Carlo simulations. These methods solve for statistical model parameters using historical data and combine them with random sampling to obtain renewable energy output scenarios. However, the effectiveness of scenario generation methods based on random sampling heavily relies on the proposed historical statistical indicators, and separate stochastic models are needed for wind power and solar power.

[0003] With the rapid development of artificial intelligence technology, using black-box models to describe the randomness of new energy output can reduce the difficulties in modeling to some extent, thus giving rise to modeling methods based on neural networks. The unsupervised learning method based on Restricted Boltzmann Machines (RBMs) can fit the nonlinear probability distribution of variables through Gibbs alternating sampling. Based on given sample training data, it generates new data that conforms to the probability distribution, making it suitable for generating medium- and long-term new energy output scenarios. Summary of the Invention

[0004] To overcome the shortcomings of the existing technology, this invention proposes a method for generating new energy output scenarios based on deep belief networks, comprising:

[0005] Step 1: Input the acquired historical power output data of new energy sources into a deep belief network based on a restricted Boltzmann mechanism, and initialize the network learning rate;

[0006] Step 2: Initialize the network parameters of the deep belief network, perform forward training on the historical output data of the new energy source, and superimpose the learned feature data onto the historical output data of the new energy source to generate new energy output training data;

[0007] Step 3: Calculate the Markov cumulative state transition probability matrix based on the power generation parameters of the new energy historical output data as feature data labels. Input the new energy output training data and feature data labels into the logistic regression layer of the deep belief network. Based on the network learning rate, use the gradient descent method to correct the feature data in the new energy output training data and output the new energy output training result data.

[0008] Step 4: Calculate the network error between the historical power output data of the new energy source and the training result data of the new energy source. Compare the network error with a preset network error threshold. When the network error is greater than the network error threshold, update the network learning rate and jump to step 2. When the network error is less than the network error threshold, output the training result data of the new energy source as the new energy source output scenario.

[0009] The deep belief network consists of a predetermined number of restricted Boltzmann machines and a single logistic regression layer.

[0010] Preferably, the initialization of the deep belief network parameters involves forward training the historical output data of the new energy source, superimposing the learned feature data onto the historical output data of the new energy source to generate new energy output training data, including:

[0011] The connection weights, explicit biases, and hidden biases of the restricted Boltzmann machine of a preset number of layers are initialized in advance;

[0012] The historical power output data of the new energy source is normalized to generate a normalized sequence of historical power output of the new energy source.

[0013] An unsupervised learning method is used to perform feature data mapping on the restricted Boltzmann machine layer by layer from bottom to top, so as to obtain feature indicators of the normalized sequence of historical power output of new energy and generate feature data.

[0014] The feature data is superimposed onto the normalized sequence of historical power output of new energy sources to generate a normalized training sequence of power output of new energy sources.

[0015] The normalized training sequence of new energy output is used as the training data for new energy output.

[0016] The feature data includes: output sequence probability distribution, monthly average output data, hourly fluctuation data, and autocorrelation coefficient; the deep belief network connects a preset number of restricted Boltzmann machines from bottom to top, and connects the top layer of restricted Boltzmann machine to the logistic regression layer. The restricted Boltzmann machine consists of explicit and hidden layers.

[0017] Preferably, the Markov cumulative state transition probability matrix calculated based on the power generation parameters of the new energy historical output data is used as the feature data label, including:

[0018] Based on the power generation state and power generation state frequency of the normalized sequence of the historical output of the new energy source, calculate the Markov state transition probability matrix.

[0019] Based on the Markov state transition probability matrix, the Markov cumulative state transition probability matrix is ​​obtained;

[0020] The Markov cumulative state transition probability matrix is ​​output as feature data labels;

[0021] The power generation status is determined by the power output status level of new energy sources.

[0022] Preferably, the step of inputting the new energy output training data and feature data labels into the logistic regression layer of the deep belief network, and using gradient descent based on the network learning rate to correct the feature data in the new energy output training data, and outputting the new energy output training result data, includes:

[0023] The normalized training sequence of renewable energy output and feature data labels in the top-level restricted Boltzmann machine are input into the logistic regression layer;

[0024] Gradient calculations are performed on each layer of the restricted Boltzmann machine from top to bottom using a backpropagation network and the gradient descent method.

[0025] By combining the gradient with the network learning rate, the changes in the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are calculated, and the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are updated according to the changes.

[0026] The deep belief network is adjusted by updating the connection weights, explicit biases, and hidden biases of each layer of the restricted Boltzmann machine, and the normalized training result sequence of new energy output is output from the bottom layer restricted Boltzmann machine.

[0027] Preferably, the changes in connection weights, explicit biases, and implicit biases of each layer of the restricted Boltzmann machine are calculated using the following formula:

[0028]

[0029] Wherein, ΔW ij Let Δa be the change in connection weights between layer i and layer j. i Let Δb be the change in bias of the i-th layer explicit element. j Let ε be the change in the bias of the hidden elements in the j-th layer, and ε be the network learning rate. For explicit metadata sampled in layer i at the 0th time, For the hidden metadata sampled in layer j for the 0th time, For the explicit data sampled in layer i at the k-th time, Let be the hidden metadata sampled in the kth time at layer j.

[0030] Preferably, the Markov cumulative state transition probability matrix is ​​calculated using the following formula:

[0031]

[0032] Among them, P′ Q Let p′ be the Markov cumulative state transition probability matrix, N be the output state level of the new energy source, and p′ be the cumulative state transition probability matrix. ij Let i be the probability of the cumulative output of the new energy source transitioning from state i to state j, where i,j = 1, 2, ..., N.

[0033] Based on the same inventive concept, this invention also provides a new energy output scenario generation system based on deep belief networks, comprising:

[0034] Historical data input module, forward training module, reverse correction module, and new energy output scenario acquisition module;

[0035] The historical data input module is used to input the acquired historical output data of new energy into a deep belief network based on a restricted Boltzmann mechanism, and to initialize the network learning rate.

[0036] The forward training module is used to initialize the network parameters of the deep belief network, perform forward training on the historical output data of the new energy, and superimpose the learned feature data onto the historical output data of the new energy to generate new energy output training data.

[0037] The reverse correction module is used to calculate the Markov cumulative state transition probability matrix based on the power generation parameters of the new energy historical output data as feature data labels, input the new energy output training data and feature data labels into the logistic regression layer of the deep belief network, and use the gradient descent method based on the network learning rate to correct the feature data in the new energy output training data, and output the new energy output training result data.

[0038] The new energy output scenario acquisition module is used to calculate the network error between the historical output data of new energy and the training result data of new energy, compare the network error with a preset network error threshold, update the network learning rate when the network error is greater than the network error threshold, and jump to the forward training module; when the network error is less than the network error threshold, output the new energy output training result data as the new energy output scenario.

[0039] The deep belief network consists of a predetermined number of restricted Boltzmann machines and a single logistic regression layer.

[0040] Preferably, the forward training module is specifically used for:

[0041] The connection weights, explicit biases, and hidden biases of the restricted Boltzmann machine of a preset number of layers are initialized in advance;

[0042] The historical power output data of the new energy source is normalized to generate a normalized sequence of historical power output of the new energy source.

[0043] An unsupervised learning method is used to perform feature data mapping on the restricted Boltzmann machine layer by layer from bottom to top, so as to obtain feature indicators of the normalized sequence of historical power output of new energy and generate feature data.

[0044] The feature data is superimposed onto the normalized sequence of historical power output of new energy sources to generate a normalized training sequence of power output of new energy sources.

[0045] The normalized training sequence of new energy output is used as the training data for new energy output.

[0046] The feature data includes: output sequence probability distribution, monthly average output data, hourly fluctuation data, and autocorrelation coefficient; the deep belief network connects a preset number of restricted Boltzmann machines from bottom to top, and connects the top layer of restricted Boltzmann machine to the logistic regression layer. The restricted Boltzmann machine consists of explicit and hidden layers.

[0047] Preferably, the Markov cumulative state transition probability matrix calculated based on the power generation parameters of historical new energy output data in the reverse correction module is used as the feature data label, including:

[0048] Based on the power generation state and power generation state frequency of the normalized sequence of the historical output of the new energy source, calculate the Markov state transition probability matrix.

[0049] Based on the Markov state transition probability matrix, the Markov cumulative state transition probability matrix is ​​obtained;

[0050] The Markov cumulative state transition probability matrix is ​​output as feature data labels;

[0051] The power generation status is determined by the power output status level of new energy sources.

[0052] Preferably, in the reverse correction module, the new energy output training data and feature data labels are input into the logistic regression layer of the deep belief network, and the feature data in the new energy output training data is corrected using the gradient descent method based on the network learning rate, and the new energy output training result data is output, including:

[0053] The normalized training sequence of renewable energy output and feature data labels in the top-level restricted Boltzmann machine are input into the logistic regression layer;

[0054] Gradient calculations are performed on each layer of the restricted Boltzmann machine from top to bottom using a backpropagation network and the gradient descent method.

[0055] By combining the gradient with the network learning rate, the changes in the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are calculated, and the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are updated according to the changes.

[0056] The deep belief network is adjusted by updating the connection weights, explicit biases, and hidden biases of each layer of the restricted Boltzmann machine, and the normalized training result sequence of new energy output is output from the bottom layer restricted Boltzmann machine.

[0057] Preferably, the changes in connection weights, explicit biases, and implicit biases of each layer of the restricted Boltzmann machine in the inverse correction module are calculated using the following formula:

[0058]

[0059] Wherein, ΔW ij Let Δa be the change in connection weights between layer i and layer j. i Let Δb be the change in bias of the i-th layer explicit element. j Let ε be the change in the bias of the hidden elements in the j-th layer, and ε be the network learning rate. For explicit metadata sampled in layer i at the 0th time, For the hidden metadata sampled in layer j for the 0th time, For the explicit data sampled in layer i at the k-th time, Let be the hidden metadata sampled in the kth time at layer j.

[0060] Preferably, the Markov cumulative state transition probability matrix in the reverse correction module is calculated using the following formula:

[0061]

[0062] Among them, P′ Q Let p′ be the Markov cumulative state transition probability matrix, N be the output state level of the new energy source, and p′ be the cumulative state transition probability matrix. ijLet i be the probability of the cumulative output of new energy source transitioning from state i to state j, where i,j = 1, 2, ..., N.

[0063] Compared with the closest existing technology, the present invention has the following beneficial effects:

[0064] 1. This invention provides a method and system for generating new energy output scenarios based on deep belief networks, comprising: Step 1: Inputting acquired historical new energy output data into a deep belief network constructed based on a restricted Boltzmann mechanism, and initializing the network learning rate; Step 2: Initializing the network parameters of the deep belief network, performing forward training on the historical new energy output data, and superimposing the learned feature data onto the historical new energy output data to generate new energy output training data; Step 3: Calculating the Markov cumulative state transition probability matrix based on the power generation parameters of the historical new energy output data as feature data labels, inputting the new energy output training data and feature data labels into the logistic regression layer of the deep belief network, using the gradient descent method based on the network learning rate to correct the feature data in the new energy output training data, and outputting new energy output training result data; Step 4: Calculating the historical new energy output data and... The network error between the training results data of the new energy power output is compared with a pre-set network error threshold. When the network error is greater than the network error threshold, the network learning rate is updated and the process jumps to step two. When the network error is less than the network error threshold, the training results data of the new energy power output are output as a new energy power output scenario. The deep belief network consists of a preset number of restricted Boltzmann machines and a logistic regression layer. This invention deeply mines the various characteristics of historical new energy power output data through the restricted Boltzmann machine and logistic regression layer of the deep belief network, resulting in a more scientific and reasonable predicted power output scenario. This method avoids the need to establish a detailed random probability distribution model of new energy power output, is not limited by resources, and can be applied to the generation of multiple scenarios such as wind power and photovoltaic power output. It only requires fine-tuning of the neural training network, which alleviates the difficulties faced in modeling the generation of medium- and long-term new energy power output scenarios.

[0065] 2. Furthermore, the generation of new energy output time series scenarios in this invention needs to fit the seasonality, volatility, and time-varying characteristics of new energy output to the greatest extent possible in order to ensure the quality of the generated scenarios. The monthly average output and monthly utilization hours can usually be used as evaluation indicators of the seasonality of new energy output series, which can well characterize the seasonal characteristics of new energy output. The autocorrelation coefficient describes the correlation between new energy output and itself under a certain time delay. The autocorrelation coefficient of the output series generated by the proposed method is close to that of the historical output series, indicating that the proposed method can well reflect the time-varying characteristics of new energy output. Attached Figure Description

[0066] Figure 1A schematic diagram of the process for generating new energy output scenarios based on deep belief networks provided by this invention;

[0067] Figure 2 A schematic diagram of a deep belief network model for a new energy output scenario generation method based on deep belief networks provided by the present invention;

[0068] Figure 3 A schematic diagram showing the comparison of historical wind power output and production output sequence of a wind farm over a period of 1200 hours, provided by the present invention.

[0069] Figure 4 This is a schematic diagram comparing the historical photovoltaic output and production output sequence of a photovoltaic power station over a period of 1200 hours, provided by the present invention.

[0070] Figure 5 The wind power output probability distribution diagram in the wind power output sequence index provided by this invention;

[0071] Figure 6 A schematic diagram of the normalized monthly average power output in the wind power output sequence index provided by this invention;

[0072] Figure 7 The probability distribution of the maximum fluctuation over 4 hours in the wind power output sequence index provided by this invention;

[0073] Figure 8 A schematic diagram of the autocorrelation coefficient in the wind power output sequence index provided by this invention;

[0074] Figure 9 This is a probability distribution diagram of photovoltaic output in the photovoltaic output sequence index provided by the present invention;

[0075] Figure 10 A schematic diagram of the normalized monthly average output in the photovoltaic output sequence index provided by this invention;

[0076] Figure 11 The maximum fluctuation probability distribution diagram of the photovoltaic power output sequence index provided by this invention is shown in the 4-hour graph.

[0077] Figure 12 This is a schematic diagram of the autocorrelation coefficient in the photovoltaic power output sequence index provided by the present invention;

[0078] Figure 13 This is a schematic diagram of a new energy output scenario generation system based on deep belief networks, provided by the present invention. Detailed Implementation

[0079] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0080] Example 1:

[0081] This invention provides a method for generating new energy output scenarios based on deep belief networks, the flowchart of which is shown below. Figure 1 As shown, it includes:

[0082] Step 1: Input the acquired historical power output data of new energy sources into a deep belief network based on a restricted Boltzmann mechanism, and initialize the network learning rate;

[0083] Step 2: Initialize the network parameters of the deep belief network, perform forward training on the historical output data of the new energy source, and superimpose the learned feature data onto the historical output data of the new energy source to generate new energy output training data;

[0084] Step 3: Calculate the Markov cumulative state transition probability matrix based on the power generation parameters of the new energy historical output data as feature data labels. Input the new energy output training data and feature data labels into the logistic regression layer of the deep belief network. Based on the network learning rate, use the gradient descent method to correct the feature data in the new energy output training data and output the new energy output training result data.

[0085] Step 4: Calculate the network error between the historical power output data of the new energy source and the training result data of the new energy source. Compare the network error with a preset network error threshold. When the network error is greater than the network error threshold, update the network learning rate and jump to step 2. When the network error is less than the network error threshold, output the training result data of the new energy source as the new energy source output scenario.

[0086] The deep belief network consists of a predetermined number of restricted Boltzmann machines and a single logistic regression layer.

[0087] Specifically, step one includes:

[0088] This invention reads historical power output data from new energy sources such as wind farms or photovoltaic power plants, and selects a preset number of restricted Boltzmann machines to construct a deep belief network. The historical power output data of new energy sources is input into the deep belief network, and the network learning rate is initialized. By selecting a suitable number of restricted Boltzmann machines to construct a deep belief network, this invention avoids inaccurate and unreasonable training data caused by too many or too few layers of restricted Boltzmann machines.

[0089] Specifically, step two includes:

[0090] like Figure 2 As shown, the deep belief network connects a predetermined number of Restricted Boltzmann Machines (RBMs) from bottom to top, and connects the top-level RBM to the logistic regression layer. The RBM consists of explicit and hidden layers, and the random neurons in each layer are represented by explicit neurons v. i Hidden element hi This indicates that the explicit element bias of each layer is a. i The hidden element bias is b i And a∈R n b∈R m W represents the connection weights of the deep belief network, W∈R n×m The random neurons in the two layers are in binary states, with a value of 1 when activated and 0 when inactive. There are connections between neurons in the Restricted Boltzmann Machine (RBM) layers, but no connections between neurons within the same layer. By training the hidden units, the correlation of related data manifested in the visible layer can be captured. The connection weights, explicit biases, and hidden unit biases of a predetermined number of RBM layers are pre-initialized. Historical power output data of new energy sources is normalized to generate a normalized sequence of historical power output. Training begins from the bottommost RBM layer and proceeds layer by layer. An unsupervised learning method is used to train the normalized historical power output sequence of new energy sources on the restricted Boltzmann machine (RBM) at each layer. After training at each layer, the characteristic indicators of the normalized historical power output sequence of new energy sources can be learned, and feature data can be generated and superimposed on the input normalized historical power output sequence of new energy sources to generate a normalized power output training sequence. During the unsupervised layer-by-layer forward training of the RBM layers, Gibbs sampling is performed alternately on the probability distribution calculation formulas of explicit and implicit elements of the RBM at each layer, and the implicit element h is extracted from it. j and the visible element v i The probability distribution formulas for the explicit element and the implicit element are as follows:

[0091]

[0092]

[0093] Where P(v|h,θ) is the probability distribution of the explicit elements, P(h|v,θ) is the probability distribution of the implicit elements, v is the explicit element, h is the implicit element, and a is the implicit element. i For the explicit element bias of the i-th layer, b j Let θ be the hidden element bias of the j-th layer, W be the connection weight of the deep belief network, θ = {a, b, W} represent the set of network parameters of the restricted Boltzmann machine, i be the number of explicit layers, j be the number of hidden layers, n be the total number of explicit layers, and m be the total number of hidden layers. Let T be the weight of the j-th layer, and the superscript T indicates the transpose of the variable.

[0094] Since the probability distribution P(v|h,θ) in a restricted Boltzmann machine is difficult to solve directly, during training, alternating Gibbs sampling is performed using the probability distribution formulas for explicit and implicit neurons to obtain the conditional probability distributions of different neurons. The specific method is as follows: h0 ~P(h|v 0 ,θ), v 1 ~P(v|h 0 ,θ),h 1 ~P(h|v 1 ,θ), …,v k+1 ~P(v|h k ,θ). The distribution of the normalized historical output sequence of new energy sources is characterized by feature data including the probability distribution of the output sequence, monthly average output data, hourly fluctuation data, and autocorrelation coefficient. This invention uses a constructed deep belief network for forward training. In the neural network of the restricted Boltzmann machine at each layer, feature data is obtained through automatic training of neurons and superimposed on the normalized historical output sequence of new energy sources to generate a normalized training sequence of new energy output. This method can fit the probability distribution of complex random variables well, deeply explore the characteristics of various aspects of the basic data, and the feature data obtained through mining can better reflect the feature indicators of the normalized historical output sequence of new energy sources, making the subsequently generated new energy output scenarios more scientific and reasonable.

[0095] Specifically, step three includes:

[0096] Deep belief networks require training of the neural network based on feature data labels. The cumulative state transition matrix of renewable energy output is used as the data training label. To reveal the transition patterns of renewable energy power generation states, a Markov chain is used to describe the state of renewable energy output at each time step. First, let Q(t) represent the output state of the renewable energy at each time step. The renewable energy output states are divided into N levels, meaning Q(t) has N states from "state 1" to "state N". Based on the power generation state statistics obtained from the normalized sequence of renewable energy output, the power generation states and their frequencies are obtained, and the Markov state transition probability matrix P is calculated. Q This is used to describe the transition probability of states at adjacent time steps, and its specific form is as follows:

[0097]

[0098] Among them, P Q Let p be the Markov state transition probability matrix. ij The probability of transitioning from state i to state j for new energy output state, where i,j=1,2,...,N, p ij The specific form is as follows:

[0099] p ij =P r (Q(t+Δt)=j|Q(t=i)

[0100] Where, p ijLet i be the probability of the new energy output state from state i to state j, i,j=1,2,...,N, Q(t+Δt) is the output state at time t+Δt, and Q(t) is the output state at time t.

[0101] According to P Q We can further obtain the Markov cumulative state transition probability matrix P′. Q :

[0102]

[0103] Among them, P′ Q Let p′ be the Markov cumulative state transition probability matrix, N be the output state level of the new energy source, and p′ be the cumulative state transition probability matrix. ij Let p′ be the probability of transitioning from state i to state j for the cumulative output of new energy, where i,j=1,2,...,N. ij The formula for calculation is:

[0104]

[0105] Where, p im The probability of transitioning from state i to state m for new energy power output, where m = 1, 2, ..., j;

[0106] The Markov cumulative state transition probability matrix is ​​output as the feature data label. The normalized training sequence of new energy output from the top-level Restricted Boltzmann Machine and the feature data label are input into the logistic regression layer. Through the backpropagation network, the gradient descent method is used to calculate the gradient of each layer of the Restricted Boltzmann Machine from top to bottom. The gradient is defined by the following formula:

[0107]

[0108]

[0109]

[0110] in, Let be the weight gradient from layer i to layer j. Let be the explicit element bias gradient of the i-th layer. Let v be the hidden element bias gradient of the j-th layer, v be the explicit element, and h be the hidden element. i h is the explicit element of the i-th layer. j Let θ = {a, b, W} be the hidden elements in the j-th layer, where a is the explicit bias, b is the hidden bias, and W is the connection weight of the deep belief network. i For the explicit element bias of the i-th layer, b j For the hidden element bias of the j-th layer, W ij Let be the connection weight from layer i to layer j.

[0111] Since the probability distribution P(v|v,θ) in a Restricted Boltzmann Machine (RBM) is difficult to solve directly, the gradient definition formula is also difficult to calculate directly. Therefore, an approximate gradient is used to replace the actual gradient for subsequent calculations. By combining the approximate gradient with the network learning rate, the changes in connection weights, explicit biases, and hidden biases of each layer of the RBM are calculated. The formulas for calculating these changes are as follows:

[0112]

[0113] Wherein, ΔW ij Let Δa be the change in connection weights between layer i and layer j. i Let Δb be the change in bias of the i-th layer explicit element. j Let ε be the change in the bias of the hidden elements in the j-th layer, and ε be the network learning rate. For explicit metadata sampled in layer i at the 0th time, For the hidden metadata sampled in layer j for the 0th time, For the explicit data sampled in layer i at the k-th time, Let be the hidden metadata sampled in the kth time at layer j.

[0114] The connection weights, explicit biases, and implicit biases of each layer of the restricted Boltzmann machine are updated based on the changes. The formulas for calculating the updated connection weights, explicit biases, and implicit biases are as follows:

[0115]

[0116] Wherein, ΔW ij Let Δa be the change in connection weights between layer i and layer j. i Let Δb be the change in bias of the i-th layer explicit element. j Let a be the change in the bias of the hidden elements in the j-th layer. i For the explicit element bias of the i-th layer, b j For the hidden element bias of the j-th layer, W ij Let be the connection weight from layer i to layer j.

[0117] The deep belief network is adjusted by updating the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine (RBM), and a normalized training result sequence of new energy output is output from the bottom-level RBM. This invention updates the network parameters of the deep belief network by calculating the changes in the connection weights, explicit biases, and hidden biases of each RBM, reducing network parameter errors and making the output results more accurate. When calculating gradients, since the gradient definition is difficult to calculate, an approximate gradient is used instead, improving the calculation speed and facilitating faster adjustment of the network parameters of the entire deep belief neural network.

[0118] Specifically, step four includes:

[0119] The network error is calculated between the normalized training result sequence of new energy output obtained after one round of deep belief network training and the initially input normalized historical output sequence of new energy. The network error is compared with a pre-set network error threshold. If the network error is greater than the threshold, the network learning rate is updated, and the process proceeds to step two. If the network error is less than the threshold, the deep belief network training is successful, and the neuron biases and weights are stored. The normalized training result sequence of new energy output is used as the new energy output training result data and is output as a new energy output scenario. This invention achieves one training cycle of the deep belief network through forward training and backward adjustment. By comparing the network error with the network error threshold, it determines whether retraining is needed. When the network error meets the standard, a new energy output scenario is output. This new energy output scenario is a medium- to long-term new energy output sequence scenario. By training the deep belief network, various medium- to long-term new energy output sequence scenarios can be generated, reducing the difficulty of modeling using existing technologies such as autoregressive moving average models, fluctuation characteristic modeling methods, and Markov-Monte Carlo simulation methods.

[0120] Example 2:

[0121] The proposed method was tested at a wind farm and a photovoltaic power station, using 8760 hours of historical output data from across the year to train the deep belief network. (Attached) Figure 3 This is a schematic diagram comparing the historical wind power output and production output of a wind farm over a 1200-hour period. (Attached) Figure 4 This is a schematic diagram comparing the historical photovoltaic output and production output sequence of a photovoltaic power station over a period of 1200 hours.

[0122] Generating time series scenarios for renewable energy output needs to fit the seasonality, volatility, and time-varying characteristics of renewable energy output to the greatest extent possible to ensure the quality of the generated scenarios. Typically, monthly average output and monthly utilization hours are used as evaluation indicators for the seasonality of the renewable energy output series; the 4-hour fluctuation probability distribution is used as an evaluation indicator for the series volatility; and the autocorrelation coefficient is used as an evaluation indicator for the series time-varying characteristics. (Appendix) Figure 5 Appendix Figure 6 Appendix Figure 7 and attached Figure 8 This diagram illustrates the probability distribution of wind power output, normalized monthly average output, probability distribution of 4-hour maximum fluctuation, and autocorrelation coefficient in the wind power output sequence indicators of a wind farm; (Attached) Figure 9 Appendix Figure 10 Appendix Figure 11 and attached Figure 12This diagram illustrates the probability distribution of photovoltaic (PV) output, the normalized monthly average output, the 4-hour maximum fluctuation probability distribution, and the autocorrelation coefficient among the PV output sequence indicators of a photovoltaic power plant. As can be seen from the diagram, the probability distribution and monthly average output of the new energy power generation sequence generated by this invention are close to the historical output sequence, indicating that the method of this invention can effectively characterize the seasonal characteristics of new energy power generation. The fluctuation probability distribution of the output sequence generated by this invention is very close to the historical output sequence, indicating that the proposed method can more accurately characterize the fluctuation characteristics of new energy power generation. The autocorrelation coefficient describes the correlation between new energy power generation and itself under a certain time delay. The autocorrelation coefficient of the output sequence generated by the proposed method is close to that of the historical output sequence, indicating that this invention can effectively reflect the time-varying characteristics of new energy power generation.

[0123] Example 3:

[0124] This invention provides a new energy output scenario generation system based on deep belief networks, the flowchart of which is shown below. Figure 13 As shown, it includes:

[0125] Historical data input module, forward training module, reverse correction module, and new energy output scenario acquisition module;

[0126] The historical data input module is used to input the acquired historical output data of new energy into a deep belief network based on a restricted Boltzmann mechanism, and to initialize the network learning rate.

[0127] The forward training module is used to initialize the network parameters of the deep belief network, perform forward training on the historical output data of the new energy, and superimpose the learned feature data onto the historical output data of the new energy to generate new energy output training data.

[0128] The reverse correction module is used to calculate the Markov cumulative state transition probability matrix based on the power generation parameters of the new energy historical output data as feature data labels, input the new energy output training data and feature data labels into the logistic regression layer of the deep belief network, and use the gradient descent method based on the network learning rate to correct the feature data in the new energy output training data, and output the new energy output training result data.

[0129] The new energy output scenario acquisition module is used to calculate the network error between the historical output data of new energy and the training result data of new energy, compare the network error with a preset network error threshold, update the network learning rate when the network error is greater than the network error threshold, and jump to the forward training module; when the network error is less than the network error threshold, output the new energy output training result data as the new energy output scenario.

[0130] The deep belief network consists of a predetermined number of restricted Boltzmann machines and a single logistic regression layer.

[0131] Specifically, the forward training module is used for:

[0132] The connection weights, explicit biases, and hidden biases of the restricted Boltzmann machine of a preset number of layers are initialized in advance;

[0133] The historical power output data of the new energy source is normalized to generate a normalized sequence of historical power output of the new energy source.

[0134] An unsupervised learning method is used to perform feature data mapping on the restricted Boltzmann machine layer by layer from bottom to top, so as to obtain feature indicators of the normalized sequence of historical power output of new energy and generate feature data.

[0135] The feature data is superimposed onto the normalized sequence of historical power output of new energy sources to generate a normalized training sequence of power output of new energy sources.

[0136] The normalized training sequence of new energy output is used as the training data for new energy output.

[0137] The feature data includes: output sequence probability distribution, monthly average output data, hourly fluctuation data, and autocorrelation coefficient; the deep belief network connects a preset number of restricted Boltzmann machines from bottom to top, and connects the top layer of restricted Boltzmann machine to the logistic regression layer. The restricted Boltzmann machine consists of explicit and hidden layers.

[0138] The Markov cumulative state transition probability matrix calculated based on the power generation parameters of historical new energy output data in the reverse correction module serves as the feature data label, including:

[0139] Based on the power generation state and power generation state frequency of the normalized sequence of the historical output of the new energy source, calculate the Markov state transition probability matrix.

[0140] Based on the Markov state transition probability matrix, the Markov cumulative state transition probability matrix is ​​obtained;

[0141] The Markov cumulative state transition probability matrix is ​​output as feature data labels;

[0142] The power generation status is determined by the power output status level of new energy sources.

[0143] Specifically, the reverse correction module inputs the new energy output training data and feature data labels into the logistic regression layer of the deep belief network, and uses a gradient descent method based on the network learning rate to correct the feature data in the new energy output training data, outputting new energy output training result data, including:

[0144] The normalized training sequence of renewable energy output and feature data labels in the top-level restricted Boltzmann machine are input into the logistic regression layer;

[0145] Gradient calculations are performed on each layer of the restricted Boltzmann machine from top to bottom using a backpropagation network and the gradient descent method.

[0146] By combining the gradient with the network learning rate, the changes in the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are calculated, and the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are updated according to the changes.

[0147] The deep belief network is adjusted by updating the connection weights, explicit biases, and hidden biases of each layer of the restricted Boltzmann machine, and the normalized training result sequence of new energy output is output from the bottom layer restricted Boltzmann machine.

[0148] Preferably, the changes in connection weights, explicit biases, and implicit biases of each layer of the restricted Boltzmann machine in the inverse correction module are calculated using the following formula:

[0149]

[0150] Wherein, ΔW ij Let Δa be the change in connection weights between layer i and layer j. i Let Δb be the change in bias of the i-th layer explicit element. j Let ε be the change in the bias of the hidden elements in the j-th layer, and ε be the network learning rate. For explicit metadata sampled in layer i at the 0th time, For the hidden metadata sampled in layer j for the 0th time, For the explicit data sampled in layer i at the k-th time, Let be the hidden metadata sampled in the kth time at layer j.

[0151] Preferably, the Markov cumulative state transition probability matrix in the reverse correction module is calculated using the following formula:

[0152]

[0153] Among them, P′ Q Let p′ be the Markov cumulative state transition probability matrix, N be the output state level of the new energy source, and p′ be the cumulative state transition probability matrix. ij Let i be the probability of the cumulative output of the new energy source transitioning from state i to state j, where i,j = 1, 2, ..., N.

[0154] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0155] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0156] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0157] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the scope of protection of the claims pending approval.

Claims

1. A method for generating new energy output scenarios based on deep belief networks, characterized in that, include: Step 1: Input the acquired historical power output data of new energy sources into a deep belief network based on a restricted Boltzmann mechanism, and initialize the network learning rate; Step 2: Initialize the network parameters of the deep belief network, perform forward training on the historical output data of the new energy source, and superimpose the learned feature data onto the historical output data of the new energy source to generate new energy output training data; Step 3: Calculate the Markov cumulative state transition probability matrix based on the power generation parameters of the new energy historical output data as feature data labels. Input the new energy output training data and feature data labels into the logistic regression layer of the deep belief network. Based on the network learning rate, use the gradient descent method to correct the feature data in the new energy output training data and output the new energy output training result data. Step 4: Calculate the network error between the historical power output data of the new energy source and the training result data of the new energy source. Compare the network error with a preset network error threshold. When the network error is greater than the network error threshold, update the network learning rate and jump to step 2. When the network error is less than the network error threshold, output the training result data of the new energy source as the new energy source output scenario. The deep belief network consists of a predetermined number of restricted Boltzmann machines and a single logistic regression layer. The network parameters of the initialized deep belief network are used to perform forward training on the historical output data of the new energy source. The learned feature data is then superimposed onto the historical output data of the new energy source to generate new energy output training data, including: The connection weights, explicit biases, and hidden biases of the restricted Boltzmann machine of a preset number of layers are initialized in advance; The historical output data of the new energy source is normalized to generate a normalized sequence of historical output of the new energy source. An unsupervised learning method is used to perform feature data mapping on the restricted Boltzmann machine layer by layer from bottom to top, so as to obtain feature indicators of the normalized sequence of historical power output of new energy and generate feature data. The feature data is superimposed onto the normalized sequence of historical power output of new energy sources to generate a normalized training sequence of power output of new energy sources. The normalized training sequence of new energy output is used as the training data for new energy output. The feature data includes: output sequence probability distribution, monthly average output data, hourly fluctuation data, and autocorrelation coefficient; the deep belief network connects a preset number of restricted Boltzmann machines from bottom to top, and connects the top layer of restricted Boltzmann machine to the logistic regression layer. The restricted Boltzmann machine consists of explicit and hidden layers.

2. The method as described in claim 1, characterized in that, The Markov cumulative state transition probability matrix calculated based on the power generation parameters of new energy historical output data serves as the feature data label, including: Based on the power generation state and power generation state frequency of the normalized sequence of the historical output of the new energy source, calculate the Markov state transition probability matrix. Based on the Markov state transition probability matrix, obtain the Markov cumulative state transition probability matrix; The Markov cumulative state transition probability matrix is ​​output as feature data labels; The power generation status is determined by the power output status level of new energy sources.

3. The method as described in claim 1, characterized in that, The process involves inputting the new energy output training data and feature data labels into the logistic regression layer of the deep belief network, using gradient descent based on the network learning rate to correct the feature data in the new energy output training data, and outputting the new energy output training result data, including: The normalized training sequence of renewable energy output and feature data labels in the top-level restricted Boltzmann machine are input into the logistic regression layer; Gradient calculations are performed on each layer of the restricted Boltzmann machine from top to bottom using a backpropagation network and the gradient descent method. By combining the gradient with the network learning rate, the changes in the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are calculated, and the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are updated according to the changes. The deep belief network is adjusted by updating the connection weights, explicit biases, and hidden biases of each layer of the restricted Boltzmann machine, and the normalized training result sequence of new energy output is output from the bottom layer restricted Boltzmann machine.

4. The method as described in claim 3, characterized in that, The changes in connection weights, explicit biases, and hidden biases of each layer of the restricted Boltzmann machine are calculated using the following formula: in, The change in connection weights between layer i and layer j. Let be the change in bias of the i-th layer explicit element. Let be the change in the bias of the hidden elements in the j-th layer. For network learning rate, For explicit metadata sampled in layer i at the 0th time, For the hidden metadata sampled in layer j for the 0th time, For the explicit data sampled in layer i at the k-th time, Let be the hidden metadata sampled in the kth time at layer j.

5. The method as described in claim 2, characterized in that, The Markov cumulative state transition probability matrix is ​​calculated using the following formula: in, Let be the Markov cumulative state transition probability matrix. The power output status level for new energy sources Accumulated power output for new energy Transition to state The probability, .

6. A new energy output scenario generation system based on deep belief networks, characterized in that, include: Historical data input module, forward training module, reverse correction module, and new energy output scenario acquisition module; The historical data input module is used to input the acquired historical output data of new energy into a deep belief network based on a restricted Boltzmann mechanism, and initialize the network learning rate. The forward training module is used to initialize the network parameters of the deep belief network, perform forward training on the historical output data of the new energy, and superimpose the learned feature data onto the historical output data of the new energy to generate new energy output training data. The reverse correction module is used to calculate the Markov cumulative state transition probability matrix based on the power generation parameters of the new energy historical output data as feature data labels, input the new energy output training data and feature data labels into the logistic regression layer of the deep belief network, and use the gradient descent method based on the network learning rate to correct the feature data in the new energy output training data, and output the new energy output training result data. The new energy output scenario acquisition module is used to calculate the network error between the historical output data of new energy and the training result data of new energy, compare the network error with a preset network error threshold, update the network learning rate when the network error is greater than the network error threshold, and jump to the forward training module; when the network error is less than the network error threshold, output the new energy output training result data as the new energy output scenario. The deep belief network consists of a predetermined number of restricted Boltzmann machines and a single logistic regression layer. The forward training module is specifically used for: The connection weights, explicit biases, and hidden biases of the restricted Boltzmann machine of a preset number of layers are initialized in advance; The historical output data of the new energy source is normalized to generate a normalized sequence of historical output of the new energy source. An unsupervised learning method is used to perform feature data mapping on the restricted Boltzmann machine layer by layer from bottom to top, so as to obtain feature indicators of the normalized sequence of historical power output of new energy and generate feature data. The feature data is superimposed onto the normalized sequence of historical power output of new energy sources to generate a normalized training sequence of power output of new energy sources. The normalized training sequence of new energy output is used as the training data for new energy output. The feature data includes: output sequence probability distribution, monthly average output data, hourly fluctuation data, and autocorrelation coefficient; the deep belief network connects a preset number of restricted Boltzmann machines from bottom to top, and connects the top layer of restricted Boltzmann machine to the logistic regression layer. The restricted Boltzmann machine consists of explicit and hidden layers.

7. The system as described in claim 6, characterized in that, The Markov cumulative state transition probability matrix calculated based on the power generation parameters of historical new energy output data in the reverse correction module serves as the feature data label, including: Based on the power generation state and power generation state frequency of the normalized sequence of the historical output of the new energy source, calculate the Markov state transition probability matrix. Based on the Markov state transition probability matrix, obtain the Markov cumulative state transition probability matrix; The Markov cumulative state transition probability matrix is ​​output as feature data labels; The power generation status is determined by the power output status level of new energy sources.

8. The system as described in claim 6, characterized in that, The reverse correction module inputs the new energy output training data and feature data labels into the logistic regression layer of the deep belief network, and uses gradient descent based on the network learning rate to correct the feature data in the new energy output training data, outputting the new energy output training result data, including: The normalized training sequence of renewable energy output and feature data labels in the top-level restricted Boltzmann machine are input into the logistic regression layer; Gradient calculations are performed on each layer of the restricted Boltzmann machine from top to bottom using a backpropagation network and the gradient descent method. By combining the gradient with the network learning rate, the changes in the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are calculated, and the connection weights, explicit biases, and hidden biases of each layer of the Restricted Boltzmann Machine are updated according to the changes. The deep belief network is adjusted by updating the connection weights, explicit biases, and hidden biases of each layer of the restricted Boltzmann machine, and the normalized training result sequence of new energy output is output from the bottom layer restricted Boltzmann machine.

Images