New energy output prediction method and system based on wavelet decomposition sequence, and related device

Abstract

The invention provides a new energy output prediction method and system based on a wavelet decomposition sequence and a related device, and the method comprises the steps: obtaining original new energy output data, and carrying out the wavelet decomposition of the new energy output data through employing a wavelet decomposition method; inputting the decomposed new energy output data into a pre-trained combined neural network, and outputting respective prediction results by each sub-neural network in the combined neural network; and based on the combined prediction result, determining an optimal integrated prediction result by using a reinforcement learning method. According to the invention, wavelet decomposition is carried out on the real output signal of the new energy, so that more detail information in the sample can be mined; model prediction is carried out by using the combined neural network, so that the inherent defects of a single neural network can be overcome, and the advantages can be combined and played to realize collaborative optimization; and the prediction result of the combined neural network forms an RQL training set, and the prediction precision and the robustness of the convergence process are further improved through a reinforcement learning method.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy output prediction, and in particular relates to a new energy output prediction method, system and related device based on wavelet decomposition sequence. Background Art

[0002] With the advancement of the market-oriented reform of my country's power system, the role of market mechanisms in resource allocation has become increasingly important. Market risks are of vital importance to the operation of market players. However, the sources and manifestations of market risks on both the supply and demand sides are different. On the demand side, market risks mainly come from the uncertainty of user loads. Under the current two-settlement framework, the specific risk is the uncertainty of revenue caused by the deviation between the planned supply on the day before and the real-time demand during the day; on the supply side, it mainly comes from new energy, especially the output uncertainty of distributed power sources, which leads to the deviation of supply on the day before. Therefore, the output and load forecasting issues on both the supply and demand sides are the core issues that must be considered and solved in the business decisions of market players.

[0003] Whether it is photovoltaic or wind turbines, their output characteristics show fluctuations in the time domain and frequency domain, and their output is intermittent, which means that if their power is directly processed as a continuous variable for modeling, it is quite deviated from reality. The existing new energy output prediction methods are difficult to take into account the requirements of accuracy and timeliness for new energy prediction. Summary of the invention

[0004] In view of this, the present invention aims to provide a new energy output prediction method, system and related devices based on wavelet decomposition sequence. The method uses wavelets to decompose new energy output data, thereby mining useful information of the time series; constructs a neural network architecture of RNN-LSTM-GRU combination to overcome the defects of a single neural network, give play to complementary advantages, and output a training set of the combined model; the training set is connected to Q-Learning, and through reinforcement learning, the intelligent agent is trained to iterate the optimal prediction result in the model-free Markov chain process, thereby improving the prediction accuracy and the robustness of the iterative process.

[0005] In order to achieve the above technical effects, the technical solution provided by the present invention is as follows:

[0006] In a first aspect, the present invention provides a new energy output prediction method based on wavelet decomposition sequence, comprising the following steps:

[0007] Obtain original new energy output data, and use wavelet decomposition method to perform wavelet decomposition on the new energy output data;

[0008] The decomposed new energy output data is input into the pre-trained combined neural network, and each sub-neural network in the combined neural network outputs its own prediction result;

[0009] Based on the combined prediction results, the reinforcement learning method is used to determine the optimal integrated prediction result, thereby determining the final prediction result of the new energy output.

[0010] Furthermore, the wavelet decomposition method is used to perform wavelet decomposition on the new energy output data, specifically including:

[0011] Discretize the power signal of the new energy station into several sampling points;

[0012] Down-sampling an initial discrete signal sequence composed of a number of sampling points to obtain a new discrete signal sequence;

[0013] The new discrete signal sequence is decomposed by wavelet through a decomposition high-pass filter and a decomposition low-pass filter to obtain two signal sequences representing the approximate part and the detail part of the original discrete signal respectively.

[0014] Furthermore, two signal sequences are as follows:

[0015]

[0016] In the formula, a j-1 Indicates the low-pass part of the original discrete signal decomposed into the next layer output, that is, the approximate part; β j-1 It represents the wavelet part of the original discrete signal decomposed into the next layer output, that is, the detail part; D(·) represents the downsampling operation, that is, removing the odd-numbered items in the sequence; H(·) and L(·) represent the high-pass and low-pass filter convolution operations respectively; α j Represents a discrete signal sequence.

[0017] Furthermore, the input of each sub-neural network in the combined neural network for new energy output prediction includes: the decomposed new energy output data and the associated variables of the new energy output.

[0018] Furthermore, the combined neural network specifically includes: RNN recurrent neural network, LSTM long short-term memory neural network and GRU recurrent gate neural network.

[0019] Furthermore, the calculation expression of the integrated prediction results is as follows:

[0020]

[0021] In the formula, represents the integrated prediction result, n represents the number of prediction results output by the n sub-neural networks of the combined neural network, represents the optimal weight of the prediction result output by the ith sub-neural network, Represents the prediction result output by the ith sub-neural network.

[0022] Furthermore, the optimal weight determination process specifically includes:

[0023] The action set from combining the prediction results to integrating the prediction results is recorded as the optimal weight corresponding to each prediction result;

[0024] Based on the historical output data of new energy and the combined prediction results, multiple trainings are carried out using the Q function update mechanism in reinforcement learning. The expression of the Q function update mechanism is as follows:

[0025]

[0026] In the formula, the subscript t represents the number of iterations, Q t+1 and Q t They represent the Q values ​​at the t+1th step and the tth step respectively, ι is a hyperparameter, R t is the reward function, γ is the discount coefficient, π represents the action set, Ω represents the target state sequence, a represents the specific action, and s t Indicates the memory status of the current step;

[0027] The best strategy given by the record corresponds to the set of actions that maximizes the Q value, as follows:

[0028] π * (Ω t ,a∈π)=maxQ

[0029] When the training converges and the prediction accuracy reaches the set requirements, the specific weight in the largest action set at this time is returned as the optimal weight, which is recorded as:

[0030]

[0031] In the formula, is the optimal weight set, is the nth optimal weight.

[0032] In a second aspect, the present invention provides a new energy output prediction system based on wavelet decomposition sequence, comprising:

[0033] A signal decomposition unit is used to obtain the original new energy output data and perform wavelet decomposition on the new energy output data using a wavelet decomposition method;

[0034] The combined prediction unit is used to input the decomposed new energy output data into the pre-trained combined neural network, and each sub-neural network in the combined neural network outputs its own prediction result;

[0035] The integrated prediction unit is used to determine the optimal integrated prediction result based on the combined prediction result by using the reinforcement learning method, so as to determine the final prediction result of the new energy output.

[0036] Accordingly, the present invention also provides a computer device, comprising: a memory and a processor and a computer program stored in the memory. When the computer program is executed on the processor, a new energy output prediction method based on wavelet decomposition sequence as in the first aspect is implemented.

[0037] Correspondingly, the present invention also provides a computer storage medium, on which a computer program is stored. When the computer program is executed by a processor, a new energy output prediction method based on wavelet decomposition sequence as in the first aspect is implemented.

[0038] In summary, the present invention provides a method, system and related device for predicting the output of new energy based on wavelet decomposition sequence, including obtaining the original output data of new energy, performing wavelet decomposition on the output data of new energy by using wavelet decomposition method; inputting the decomposed output data of new energy into a pre-trained combined neural network, and each sub-neural network in the combined neural network outputs its own prediction result; based on the combined prediction result, using the reinforcement learning method to determine the optimal integrated prediction result, thereby determining the final prediction result of the output of new energy. The present invention performs wavelet decomposition on the real output signal of new energy, which can mine more detailed information in the sample; using the combined neural network for model prediction can not only overcome the inherent shortcomings of a single neural network, but also combine its advantages to achieve collaborative optimization; the prediction results of the combined neural network are formed into a training set of RQL, and the prediction accuracy and robustness of the convergence process are further improved by the reinforcement learning method. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 A flowchart of a new energy output prediction method based on wavelet decomposition sequence provided by an embodiment of the present invention;

[0041] Figure 2 A structural diagram of a new energy output prediction system based on wavelet decomposition sequence provided by an embodiment of the present invention;

[0042] Figure 3 A structural diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] The prior art is first further introduced below.

[0045] Regarding the output and load forecasting issues on both the supply and demand sides, in recent years, with the development of cutting-edge technologies such as artificial intelligence, new technologies such as neural networks and deep learning have been widely used in the energy field, including load forecasting on the demand side and new energy output forecasting on the supply side. However, it is necessary to note the inherent differences between load forecasting and new energy output forecasting. For load forecasting, the focus is on examining characteristic variables, because there are many factors affecting its uncertainty, including user habits, production activities, and possibly exogenous factors such as macroeconomic conditions; and for the prediction of new energy output, due to the electrical principles and physical factors of its output, the research on the internal driving factors is actually very detailed, so more attention should be paid to the useful information contained in the historical data itself. In other words, due to the different inherent driving mechanisms of their time series and characteristic variables, there must be significant differences in their technical paths for new energy output forecasting and load forecasting.

[0046] The various neural networks and machine learning methods currently in use have different advantages and disadvantages, and their scope of application is also different. Taking neural networks as an example, CNN convolutional neural networks are good at mining feature information in short- and medium-term sequences, but they perform poorly in the prediction of long-term sequences; LSTM long short-term memory networks are good at processing long-distance time series, but are not as good as CNN in mining high-dimensional features. Therefore, the current visible trend is to organically integrate various neural networks and machine learning methods, and produce complementary effects by using different methods to improve the accuracy and timeliness of load forecasting and new energy output forecasting. In addition, in the integration process of multiple methods and architectures, the differences between new energy output forecasting and load forecasting must still be considered. As mentioned earlier, unlike the high-dimensional characteristics of variables in load forecasting, the focus of new energy output forecasting is not the feature mining of independent variables, but the effective information mining of the time series of its real output data itself.

[0047] Whether it is photovoltaic or wind turbines, the characteristics of their output show fluctuations in the time domain and frequency domain, and the intermittent nature of their output means that if their power is directly processed as a continuous variable for modeling, it is relatively far from reality. In fact, the signal analysis of renewable energy output and the processing method of mining time series information in discrete signals are closer to reality. These signal analysis methods include Fourier transform and wavelet packet analysis. One disadvantage of the Fourier series is that its building blocks are periodic sine waves and cosine waves with no beginning or end. This makes this method suitable for filtering or compressing those fluctuating signals with approximate periodicity, while sine waves and cosine waves are powerless in the face of signals with significant local characteristics. Wavelet packets are suitable for simulating mutation signals - signals with significant local characteristics, so they have been used in the time domain and frequency domain of renewable energy power. After mining effective information, we should also consider integrated prediction of data. At present, the prediction of new energy output mainly includes two paths. One is to provide training and analysis architecture through combined neural networks, and then use reinforcement learning or deep learning methods to train intelligent agents; the other is to iteratively optimize basic predictors through integrated learning methods such as boosting.

[0048] The present invention proposes a new energy output prediction method, system and related devices based on wavelet decomposition sequence. Taking into account the requirements for accuracy and timeliness of new energy prediction, the present invention proposes a combined neural network as the training framework of the intelligent agent; considering that the intelligent agent needs to directly apply the prediction to various decision-making activities of market entities, the output of the combined neural network is selected to be reinforced with RQL training, thereby providing a global optimal prediction method that does not require a priori model in the Markov process.

[0049] The embodiments of the present invention are described below.

[0050] See also Figure 1 This embodiment provides a new energy output prediction method based on wavelet decomposition sequence, comprising the following steps:

[0051] S1: Obtain the original new energy output data, and use the wavelet decomposition method to perform wavelet decomposition on the new energy output data.

[0052] It should be noted that, as mentioned above, the output of photovoltaic and wind turbines has volatility in the time domain and frequency domain. Directly modeling the time period data as a continuous variable will lose a lot of useful information in the data, so it is necessary to perform signal decomposition on the output data. The present invention uses wavelet packet decomposition (WPD) to perform output on the output data. Compared with the standard wavelet transformation, wavelet decomposition can not only provide more flexible signal analysis, but also adapt to specific signal characteristics, so it is particularly useful in noise reduction and feature extraction.

[0053] S2: Input the decomposed new energy output data into the pre-trained combined neural network, and each sub-neural network in the combined neural network outputs its own prediction result.

[0054] It should be noted that, considering the characteristics of renewable energy output, its power prediction focuses more on mining the information of the time series itself, rather than considering the relationship between sample characteristics and model accuracy, because the principles of photovoltaic and wind power output are already quite clear. Therefore, several sub-neural networks can be used in parallel to form a hidden layer of a composite neural network, and the individual prediction results output by the three networks are used as the data set for the next step of the intelligent agent to perform RQL.

[0055] S3: Based on the combined prediction results, the reinforcement learning method is used to determine the optimal integrated prediction result, thereby determining the final prediction result of the new energy output.

[0056] It should be noted that after the combined neural network outputs a complete prediction sequence, the goal of this solution is to optimize the results of the aforementioned integrated neural network prediction by training the intelligent agent through reinforcement Q-learning (Reinforcing Q-Learning), thereby improving the prediction accuracy and the efficiency and robustness of the algorithm convergence.

[0057] RQL is a model-free reinforcement learning method, which means that the essence of its training is to set up incentive mechanisms to strengthen the agent to search for the best strategy in the Markov process, which corresponds to an action and a corresponding state value. is a result that has been given before, so it can be regarded as the state space and data set of the agent training. Then the goal of training is to determine the best method to integrate the prediction results. This goal includes two parts at the application level. One is to determine to output the integrated prediction result; the other is to improve the prediction accuracy after integration.

[0058] This embodiment proposes a new energy output prediction method based on wavelet decomposition sequence, which performs wavelet decomposition on the real output signal of new energy, and can mine more detailed information in the sample; uses combined neural networks for model prediction, which can not only overcome the inherent shortcomings of a single neural network, but also combine its advantages to achieve collaborative optimization; the prediction results of the combined neural network are formed into a training set of RQL, and the prediction accuracy and the robustness of the convergence process are further improved through reinforcement learning methods.

[0059] In one embodiment of the present invention, the specific process of wavelet decomposition of output data is as follows:

[0060] First, the power signal monitored by the new energy station is discretized into 2 t sampling points, so we get

[0061]

[0062] In the above formula, f(·) represents the approximate mapping of the information space; represents the discrete signal before wavelet decomposition, the superscript t represents the tth approximation step of the approximate mapping, and the subscript k represents the signal position in the tth step.

[0063] Define the downsampling operator, for a sequence {…,x -2 ,x -1 ,x0,x1,x2,…}, downsampling means removing the odd-numbered items in the sequence, then the downsampling sequence of the sequence can be expressed as

[0064] D x ={…,x -2 ,x0,x2,…x n-2 ,x n} (2)

[0065] The filter of wavelet decomposition uses a decomposition high-pass filter and a decomposition low-pass filter, denoted as h and l respectively. Define two convolution operators H and L, the expressions are as follows

[0066]

[0067] In the above formula, H(x) represents the high-pass filter convolution of sequence x; correspondingly, L(x) represents the low-pass filter convolution of sequence x; * represents the convolution symbol.

[0068] Then the α of the discrete signal of the j-th iteration of the original sequence is j The wavelet decomposition outputs can be expressed as

[0069]

[0070] In the above formula, a j-1 It represents the low-pass part of the original discrete signal decomposed into the next layer output, which can represent the approximate part of the original information; β j-1 The wavelet part (detail part) of the original discrete signal decomposed into the next layer output can represent the hidden information in the original information. By decomposing it level by level, to level 0, the iterative process is completed.

[0071] Since the discrete filter and downsampling operators do not depend on j, the storage requirement is small, and since the convolution operation is not time-consuming, the entire iterative process is fast and efficient.

[0072] In one embodiment of the present invention, the power signal monitored by the new energy station is converted into two sequences through wavelet decomposition, which can be entered into a combined neural network for prediction together with variable sequences such as weather and radiation.

[0073] In one embodiment of the present invention, the combined neural network may include an RNN recurrent neural network, an LSTM long short-term memory neural network, and a GRU recurrent gate neural network. The following is a detailed introduction to each module in the combined neural network of this embodiment.

[0074] 1.RNN module

[0075] RNN is a special neural network structure that not only considers the input of the previous moment, but also gives the network a "memory" function of the previous content. RNN is called a recurrent neural network because the current output of the sequence is also related to the previous output. The specific manifestation is that the network will remember the previous information and apply it to the calculation of the current output, that is, the nodes between the hidden layers are no longer disconnected but connected, and the input of the hidden layer includes not only the output of the input layer but also the output of the hidden layer at the previous moment.

[0076] A basic RNN structure consists of an input layer, a hidden layer, and an output layer. Let the input sequence of step t be x t , the hidden layer processes two states, namely the memory state of the previous step and the state processed by the input layer of this time step. The memory state of the current step depends on the state of the hidden layer at the current step, and the output of the output layer depends on the memory state of the current step. Starting from the initial state, the cycle continues, and the states of the three are interrelated until convergence, and the hidden layer outputs the final prediction result.

[0077] Assuming that the program converges at step t, the hidden layer state at the current step can be expressed as

[0078] h t =Ux t +Ws t-1 (5)

[0079] In the above formula, h t represents the hidden layer state at step t; U represents the sample weight matrix input at this step; W represents the weight matrix of the input layer; s t-1 Indicates the memory status of the previous step.

[0080] Then the memory state of the current step can be expressed as

[0081] s t =ReLU(h t ) (6) In the above formula, ReLU is the activation function. Correspondingly, the output at the current step can be expressed as

[0082] o t =softmax(V·s t ) (7)

[0083] In the above formula, o t represents the output of RNN at step t; softmax is the activation function; V represents the weight matrix of the output layer. In the above formulas, U, W and V are set at the initial step and remain unchanged in the subsequent iterations, that is, the weights are shared in all iteration steps.

[0084] The sequence obtained by wavelet decomposition is input into RNN. In the convergence step, two prediction result sequences of RNN can be obtained. Reconstructing the two sequences can obtain the RNN single prediction result, which can be recorded as

[0085] 2. LSTM Module

[0086] The main structure of a single LSTM includes memory cells, input gates, output gates, and forget gates. The memory cell is the core of LSTM, responsible for saving important information and passing it to the subsequent network layers. The role of the memory cell is similar to the memory of the brain. Its characteristic is that it can save historical information and apply it to the current prediction. The neuron is updated at each time step, and then the input gate, forget gate, and output gate control how to update and output this value at each time step. The input gate determines whether the current input information is written to the memory cell, that is, it can control the impact of the input information on the memory cell. The forget gate determines which information the memory cell should retain and forget other unimportant information; the output gate determines which information in the memory cell should be output to the subsequent neural network.

[0087] These four parts calculate the dot product of the weight matrix and the input signal, and calculate the output value of each gate through the activation function (usually the Sigmoid function), and then multiply it by the state value of the memory cell for the final calculation. The input gate determines how much information to write into the memory cell by multiplying its output value by the current input signal, the forget gate determines how much information to clear by multiplying its output value by the current state value of the memory cell, and the output gate determines how much information to output to the subsequent network layer by multiplying its output value by the current value of the memory cell. Through such a structural design, LSTM can forget irrelevant information while retaining important information, and better handle the problem of long-term dependency.

[0088] At time t, LSTM will output a hidden layer data sequence h t , at this time the memory cell has a corresponding state data. After entering the next layer of loop, at the input end, LSTM receives the hidden layer data sequence h of the previous moment t-1 、The state of memory cell at the last moment c t-1 , and the input sequence x at time t t .

[0089] The output state of the input gate at time t can be expressed as

[0090] I t =σ(W I ·[h t-1 , x t ]+d I ) (7)

[0091] In the above formula, σ represents the Sigmoid function; I t is the output of the input gate at time t; W I represents the weight matrix of the input gate; d C Represents the amount of drift between the input gate and the memory cell.

[0092] At this time, the state update value of the memory cell controlled by the input gate is

[0093]

[0094] In the above formula, represents the state update value of the memory cell determined by the input gate at time t; tanh represents the hyperbolic tangent function; W C Represents the weight matrix of memory cell state update; d C Represents the drift of cell state update.

[0095] The forget gate determines the cell state c at the previous moment t-1 Which information should be remembered in the memory? Therefore, the output of the forget gate at time t is

[0096] F t =σ(W F ·[h t-1 , x t ]+d F ) (10)

[0097] In the above formula, W F Represents the weight matrix of the forget gate; d F Represents the drift between the forget gate and the memory cell. After the output of the input gate and the forget gate is given, the state of the memory cell at time t is

[0098]

[0099] At this time, the output h of the hidden layer at time t t It is determined by the output of the output layer at time t and the state of the memory cell at time t, and its expression is:

[0100] h t =O t ⊙tanh(c t ) (12) Among them, O t Given by the following formula

[0101] O t =σ(W O ·[h t-1 , x t ]+d O ) (13)

[0102] Assuming that the algorithm converges at step t, the sequence obtained by wavelet decomposition is input into LSTM. At the convergence step, two prediction result sequences of LSTM can be obtained. Reconstructing the two sequences can obtain the LSTM single prediction result, which can be recorded as

[0103] 3. GRU module

[0104] GRU (Gate Recurrent Unit) is an improved RNN proposed to solve problems such as long-term memory and gradient in back propagation. Three gate functions are introduced in LSTM: input gate, forget gate and output gate to control input value, memory value and output value. In the GRU model, there are only two gates: update gate and reset gate. Structurally, the reset gate is actually a combination of the forget gate and output gate in LSTM.

[0105] In the tth step of the iterative update, the input sequence is still recorded as x t At this time, for the hidden layer of GRU, the input also includes the hidden layer input h of the previous step t-1 , then the state of the reset gate at the current step can be expressed as

[0106] r t =σ(W r ·[h t-1 , x t ]+d r ) (14)

[0107] In the above formula, r t Represents the state update of the reset gate at the current step; σ is still the Sigmoid function; W r Represents the weight matrix of the reset gate; d r Indicates the drift amount for resetting the gate.

[0108] Correspondingly, the state of the update gate at the current step can be expressed as

[0109] z t =σ(W z ·[h t-1 ,x t ]+d z ) (15)

[0110] In the above formula, z tIndicates the state update of the update gate at the current step; W z represents the weight matrix of the update gate; d z Indicates the drift of the update gate.

[0111] While resetting the gate to update its own state, it also outputs a hidden layer component, which is expressed as

[0112]

[0113] In the above formula, tanh represents the activation function; W h Represents the connection weight matrix of the reset gate; d h is the corresponding drift amount.

[0114] The final output of the hidden layer of GRU is

[0115]

[0116] Similarly, assuming that the algorithm converges at step t, the sequence obtained by wavelet decomposition is input into GRU. At the convergence step, two prediction result sequences of GRU can be obtained. Reconstructing the two sequences can obtain the separate prediction result of GRU, which can be recorded as

[0117] Compared with LSTM, GRU has fewer parameters and therefore converges faster, but its ability to mine key features of sample time series is not as good as LSTM. The above three models have their own advantages and disadvantages. Combining them can bring about synergistic and complementary effects. Then we get the prediction sequence output by each of the three models

[0118] In one embodiment of the present invention, a reinforcement learning method is used to determine an integrated prediction result. The reinforcement learning integrated prediction is introduced below.

[0119] After the combined neural network outputs a complete prediction sequence, the goal is to optimize the results of the aforementioned integrated neural network predictions through reinforcement Q-learning (Reinforcing Q-Learning) training agents, thereby improving the accuracy of the predictions and the efficiency and robustness of the algorithm convergence.

[0120] RQL is a model-free reinforcement learning method, which means that the essence of its training is to set up incentive mechanisms to strengthen the agent to search for the best strategy in the Markov process, which corresponds to an action and a corresponding state value. is a result that has been given before, so it can be regarded as the state space and data set of the agent training. Then the goal of training is to determine the best method to integrate the prediction results. This goal includes two parts at the application level. One is to determine to output the integrated prediction result; the other is to improve the prediction accuracy after integration.

[0121] Therefore, we can first set the target state sequence

[0122]

[0123] In the above formula, Indicates the value returned by the agent when determining the optimal action The weight of .

[0124] Then for the target state, the action set anchored by the agent can be expressed as

[0125]

[0126] In the above formula, the subscript i represents the object to which the action is anchored; t represents the iteration step. The weight of .

[0127] Since the goal of the intelligent agent is to improve prediction accuracy, the loss function can be selected in the form of mean square error. The expression of the loss function is:

[0128]

[0129] In the above formula, y i Represents the real power sample data of the i-th new energy source; Represents the predicted value given by RQL at the current iteration; n represents the total number of samples. For the incentive mechanism of the agent, the ∈-greedy strategy can be adopted. The reward and punishment of the agent are directly linked to the prediction accuracy, without considering the sparse reward problem. Then the reward function is set to

[0130]

[0131] In the above formula, χ t is a dummy variable whose value is

[0132]

[0133] The meaning of the above two equations is that at step t, if the prediction accuracy of the agent is improved, it will be rewarded, otherwise, there will be no reward. Based on the above settings, the update mechanism of the Q function can be expressed as

[0134]

[0135] In the above formula, ι is a hyperparameter, which indicates the learning efficiency; γ is the discount coefficient; and π indicates the action set. After multiple trainings, the best strategy given by the agent corresponds to the action set that maximizes the Q value, so we have

[0136] π * (Ω t ,a∈π)=maxQ (24)

[0137] When the model converges and the prediction accuracy reaches the requirement, the agent returns π * (Ω t ,a∈π) in Ω t The specific weight is recorded as The optimized final prediction result is obtained, which can be expressed as

[0138]

[0139] Due to the adoption of the above scheme, the present invention has the following advantages:

[0140] The wavelet decomposition of the real output signal of new energy can dig out more detailed information in the sample; the use of RNN, LSTM and GRU combined neural network for model prediction can not only overcome the inherent shortcomings of a single neural network, but also combine their advantages to achieve collaborative optimization; the prediction results of the combined neural network are formed into the RQL training set, and the prediction accuracy and robustness of the convergence process are further improved through reinforcement learning methods.

[0141] Based on the same inventive concept, the embodiment of the present application also provides a new energy output prediction system based on wavelet decomposition sequence of a new energy output prediction method based on wavelet decomposition sequence. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiment of the new energy output prediction system based on wavelet decomposition sequence provided below can be referred to the limitations of the new energy output prediction method based on wavelet decomposition sequence above, which will not be repeated here.

[0142] See also Figure 2 This embodiment provides a new energy output prediction system based on wavelet decomposition sequence, including:

[0143] A signal decomposition unit is used to obtain the original new energy output data and perform wavelet decomposition on the new energy output data using a wavelet decomposition method;

[0144] The combined prediction unit is used to input the decomposed new energy output data into the pre-trained combined neural network, and each sub-neural network in the combined neural network outputs its own prediction result;

[0145] The integrated prediction unit is used to determine the optimal integrated prediction result based on the combined prediction result by using the reinforcement learning method, so as to determine the final prediction result of the new energy output.

[0146] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0147] Reference Figure 3 The embodiment of the present invention further provides a computer device 3, comprising: a memory 302 and a processor 301 and a computer program 303 stored in the memory 302. When the computer program 303 is executed on the processor 301, a new energy output prediction method based on wavelet decomposition sequence as described in any one of the above methods is implemented.

[0148] The computer device 3 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will appreciate that Figure 3 It is only an example of computer device 3 and does not constitute a limitation on computer device 3. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components, for example, it may also include input and output devices, network access devices, etc.

[0149] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0150] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may also be an external storage device of the computer device 3, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the computer device 3. Further, the memory 302 may also include both an internal storage unit of the computer device 3 and an external storage device. The memory 302 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program, etc. The memory 302 may also be used to temporarily store data that has been output or is to be output.

[0151] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting new energy output based on wavelet decomposition sequence as described in any one of the above methods is implemented.

[0152] In this embodiment, if the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the camera / terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. For example, a USB flash drive, a mobile hard disk, a disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0153] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0154] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0155] In the embodiments disclosed in the present application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A new energy output prediction method based on wavelet decomposition sequence, characterized in that: The steps include: Acquire original new energy output data, and perform wavelet decomposition on the new energy output data using a wavelet decomposition method; Inputting the decomposed new energy output data into a pre-trained combined neural network, and each sub-neural network in the combined neural network outputs its own prediction result; Based on the combined prediction results, the reinforcement learning method is used to determine the optimal integrated prediction result, thereby determining the final prediction result of the new energy output.

2. The new energy output prediction method based on wavelet decomposition sequence according to claim 1 is characterized in that: The wavelet decomposition method is used to perform wavelet decomposition on the new energy output data, specifically including: Discretize the power signal of the new energy station into several sampling points; Down-sampling the initial discrete signal sequence composed of the plurality of sampling points to obtain a new discrete signal sequence; The new discrete signal sequence is subjected to wavelet decomposition through a decomposition high-pass filter and a decomposition low-pass filter respectively, and the output is obtained to obtain two signal sequences representing the approximate part and the detail part of the original discrete signal respectively.

3. The new energy output prediction method based on wavelet decomposition sequence according to claim 2 is characterized in that: The two signal sequences are as follows: In the formula, a j-1 Represents the low-pass part of the original discrete signal decomposed into the next layer output, that is, the approximate part; β j-1 represents the wavelet part of the original discrete signal decomposed into the next layer output, that is, the detail part; D(·) represents the downsampling operation, that is, removing the odd-numbered items in the sequence; H(·) and L(·) represent the high-pass and low-pass filter convolution operations respectively; α j Represents a discrete signal sequence.

4. The new energy output prediction method based on wavelet decomposition sequence according to claim 1 is characterized in that: The input of each sub-neural network in the combined neural network for predicting the output of new energy includes: the decomposed output data of the new energy and the associated variables of the output of new energy.

5. The new energy output prediction method based on wavelet decomposition sequence according to claim 1 is characterized in that: The combined neural network specifically includes: RNN recurrent neural network, LSTM long short-term memory neural network and GRU recurrent gate neural network.

6. The new energy output prediction method based on wavelet decomposition sequence according to claim 1 is characterized in that: The calculation expression of the integrated prediction result is as follows: In the formula, represents the integrated prediction result, n represents the number of prediction results output by the n sub-neural networks of the combined neural network, represents the optimal weight of the prediction result output by the ith sub-neural network, Represents the prediction result output by the ith sub-neural network.

7. The new energy output prediction method based on wavelet decomposition sequence according to claim 6 is characterized in that: The optimal weight determination process specifically includes: The action set from the combined prediction result to the integrated prediction result is recorded as the optimal weight corresponding to each prediction result; Based on the historical output data of new energy sources and the combined prediction results, multiple trainings are performed using the Q function update mechanism in the reinforcement learning. The expression of the Q function update mechanism is as follows: In the formula, the subscript t represents the number of iterations, Q t+1 and Q t They represent the Q values ​​at the t+1th step and the tth step respectively, ι is a hyperparameter, R t is the reward function, γ is the discount coefficient, π represents the action set, Ω represents the target state sequence, a represents the specific action, and s t Indicates the memory status of the current step; The best strategy given by the record corresponds to the set of actions that maximizes the Q value, as follows: p * (Oh t ,a∈π)=max Q When the training converges and the prediction accuracy reaches the set requirements, the specific weight in the largest action set at this time is returned as the optimal weight, which is recorded as: In the formula, is the optimal weight set, is the nth optimal weight.

8. A new energy output prediction system based on wavelet decomposition sequence, characterized in that: include: A signal decomposition unit, used for obtaining original new energy output data, and performing wavelet decomposition on the new energy output data by using a wavelet decomposition method; A combined prediction unit, used for inputting the decomposed new energy output data into a pre-trained combined neural network, and each sub-neural network in the combined neural network outputs its own prediction result; The integrated prediction unit is used to determine the optimal integrated prediction result based on the combined prediction result by using the reinforcement learning method, so as to determine the final prediction result of the new energy output.

9. A computer device, characterized in that: include: A memory and a processor and a computer program stored in the memory, when the computer program is executed on the processor, implements a new energy output prediction method based on wavelet decomposition sequence as described in any one of claims 1-7.

10. A computer storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for predicting new energy output based on wavelet decomposition sequence according to any one of claims 1 to 7 is implemented.

Images