A method for training deep learning models of power systems

Abstract

A method for training a deep learning model for a power system includes the following steps: Step 1: Constructing a loss function based on the Mean Absolute Error (MAE) using an Encoder-Decoder architecture; Step 2: Using the Adam algorithm as the neuron weight update algorithm to train the parameters of each GRU neuron within the Encoder-Decoder architecture; Step 3: Adaptively finding the learning rate for each parameter using the Adam algorithm. The purpose of this invention is to solve the technical problem of low training efficiency when using a deep learning model based on Seq2Seq technology to train differential sample data, due to the fact that the unit start-up and shutdown state matrices and output state matrices of the actual power system are high-dimensional sample matrices.

Description

Technical Field

[0001] This invention belongs to the field of power system and automation research, and in particular relates to the research on unit combination decision-making methods based on deep learning intelligent algorithms. This invention is a divisional application of the invention entitled "Data-driven Intelligent Decision-making Method for Unit Combination Based on GRU and Seq2Seq Technology" (application number 2019108724540). Background Technology

[0002] Open and mature electricity markets often require an Independent System Operator (ISO) to maintain their safe and reliable operation. The Security Constrained Unit Commitment (SCUC) problem, which considers security constraints, is a crucial theoretical foundation for ISO decision-making. In recent years, the global electricity market has developed rapidly. This necessitates that ISO possess powerful computational tools to maintain market operations and formulate intelligent and precise day-ahead generation plans. Furthermore, with the widespread application of new energy technologies such as electric vehicles, intermittent energy sources, and demand-side management, the challenges faced by ISO decision-making are constantly emerging. Therefore, researching highly adaptable and accurate SCUC decision-making theories has significant theoretical and engineering implications for the development and marketization of the power industry.

[0003] Depending on the factors considered, current unit combination research can be broadly categorized into several types, including multi-objective unit combination, uncertain unit combination, and unit combination considering multiple constraints and decision variables. Overall, while current research on unit combination problems has different focuses, it generally starts with practical engineering problems, proposes mathematical models based on mechanistic studies, and then studies corresponding mathematical methods for solving them. This type of research is grounded in rigorous logical deduction and mechanistic research, driven by mathematical models and algorithms, and can therefore be called a physical model-driven unit combination decision-making method. However, because this method requires modification and reconstruction of the model itself when facing emerging new problems, its applicability may be insufficient in today's rapidly changing energy landscape and the constant emergence of theoretical challenges.

[0004] In practical engineering terms, once unit combination decision-making methods are applied, they often accumulate a large amount of structured historical data. In the long run, unit combination decisions also exhibit a certain degree of repetition. If a data-driven unit combination decision-making method could be proposed, without studying its underlying mechanisms, but instead based on deep learning methods and trained using massive amounts of historical decision data, directly constructing the mapping relationship between known inputs and decision results, and continuously refining the model through the accumulation of historical data, then the unit combination decision-making process could be endowed with self-evolution and self-learning capabilities. This data-driven decision-making method could not only greatly simplify the modeling and solving process of unit combination problems and reduce its complexity, but also address emerging theoretical problems and challenges through its self-learning. However, research in this field is currently relatively limited. Some literature addresses the temporal characteristics of unit combination sample data. For example, the paper "Research on a Data-Driven Intelligent Decision-Making Method for Unit Combination with Self-Learning Capability" first employed a recurrent neural network, namely Long Short-Term Memory (LSTM), as the core training tool to successfully construct a data-driven unit combination decision-making model and verified its self-evolutionary characteristics and adaptability to different unit combination problems. However, the method in this paper still has the following problems:

[0005] 1) Due to the excessive complexity of LSTM models, they require significant computational resources and are prone to overfitting when processing high-dimensional training samples. In contrast, the latest improved version of recurrent neural networks, the Gated Recurrent Unit (GRU), combines the input and forget gates and simplifies the memory unit, effectively reducing the model's complexity.

[0006] 2) Historical data of power systems varies greatly (e.g., load characteristics differ significantly across seasons). If a single recurrent neural network architecture is used for offline training, a single compromise mapping model will inevitably be generated when faced with such diverse historical sample data, making it difficult to guarantee the accuracy of online decision-making. To address this, this paper proposes a clustering training approach: first, the historical scheduling data is clustered, and then each cluster is trained separately to obtain multiple mapping models. During decision-making, the type of input data is first determined, and then the corresponding mapping model is used for online decision-making. While this approach partially solves the fitting accuracy problem of a single recurrent neural network architecture when facing diverse samples, the need to train multiple deep learning models significantly reduces both training and decision-making efficiency.

[0007] As an effective means of solving sequence-based problems, sequence-to-sequence (Seq2Seq) technology has been widely used in machine translation, intelligent question answering, and other fields in recent years. Unlike traditional single recurrent neural network (RNN) architectures that use a single neuron to read all input data and output the result, Seq2Seq technology uses two RNNs to form an Encoder-Decoder architecture. The Encoder reads the input sequence step-by-step according to time steps and then outputs the intermediate state C of the entire sequence. Since RNNs can record the process information of each training step, theoretically, the intermediate state C can take into account the information of the entire input sequence. In the Decoder, another RNN performs the opposite operation to the Encoder, decoding the obtained intermediate state C step-by-step to form the final output sequence. Because the intermediate state C can completely preserve the category information and pointing probabilities of both the input and output sequences, theoretically, Seq2Seq technology holds promise as a feasible approach to solving the problem that single RNN models cannot accurately train on differentially expressed sample data. However, since the dimension of the unit start-up and shutdown matrix is ​​proportional to the number of units in the system, the actual power system's unit start-up and shutdown state matrix and output state matrix are both high-dimensional sample matrices. Directly training these data using a deep learning model based on Seq2Seq technology results in low training efficiency. Therefore, it is necessary to introduce Seq2Seq technology while also researching dimensionality reduction strategies for unit start-up and shutdown sample data to further improve training efficiency. Summary of the Invention

[0008] The purpose of this invention is to solve the technical problem of low training efficiency when using a deep learning model based on Seq2Seq technology to train differential sample data, because the generator start-up and shutdown state matrix and the output state matrix of the actual power system are both high-dimensional sample matrices.

[0009] The technical solution adopted in this invention is:

[0010] A data-driven intelligent decision-making method for unit cohesion based on GRU and Seq2Seq technologies includes the following steps:

[0011] 1. Use sample encoding techniques for high-dimensional unit combination training sample matrices to compress the dimensionality of historical decision-making data for unit combination;

[0012] 2. Based on threshold recurrent networks, Seq2Seq technology is introduced to establish a composite neural network architecture for unit combination decision-making;

[0013] 3. Based on this, a deep learning model for unit combination is constructed, and a mapping model between system daily load and unit start-up and shutdown schemes is established through training with historical data;

[0014] 4. Use the generated mapping model to make unit combination decisions, obtain the unit start-up and shutdown status and unit output under the optimal power flow model, and use the obtained unit combination decision results as new historical sample data to train the deep learning model, thereby realizing continuous improvement of the model.

[0015] In step 1, when processing the high-dimensional unit combination training sample matrix, specifically, the start-stop state vector of the unit combination for each time period is encoded so that vectors with the same start-stop state are also encoded in the same way.

[0016] The start-up and shutdown status vector of each unit combination in each time period is converted into a corresponding decimal code, thereby compressing the dimension of the sample matrix.

[0017] In step 2, an Encoder-Decoder composite neural network architecture is constructed based on GRU and Seq2Seq technologies, specifically through the following steps:

[0018] 1) Transfer a historical mapping sample (P) L U G Substituting this into the Encoder-Decoder architecture, where P L For daily load data, U G For the corresponding unit start-up and shutdown scheme, the Encoder architecture will P L The data is read in steps, and the hidden layer state of the GRU neuron at time t is jointly determined by the hidden layer state of the GRU neuron at time t-1 and the daily load at time t. The specific formula is as follows:

[0019] h t =f(h) t-1 ,P Lt (1)

[0020] Where: h t h represents the hidden state of the GRU neuron at time t; t-1 P represents the hidden state of the GRU neuron at time t-1; Lt This represents the daily load input at time t;

[0021] 2) In the Encoder architecture, the hidden state h of the GRU neuron at time t. t Similar to the intermediate states in the Encoder architecture, in the Decoder architecture, the hidden state h of the GRU neuron at time k is... k Similar to the intermediate states of the Decoder architecture, the specific formula is:

[0022]

[0023] In the formula: C t Represents the intermediate state of the Encoder architecture at time t; C k This represents the intermediate state of the Encoder architecture at time k;

[0024] 3) The intermediate state output by the Encoder architecture at time T is the intermediate state C of the input sequence, with a value of C. T , representing the complete information of the input sequence, specifically as follows:

[0025] C = C T (3);

[0026] 4) Input the intermediate state C of the sequence into the Decoder architecture. The initial value C0 of the Decoder intermediate state is the same as the intermediate state C of the sequence. After inputting C0, the hidden state h of the GRU neuron at time k can be obtained. k It is jointly determined by the hidden state of the GRU neuron at time k-1 and the input of the GRU neuron at time k, and the specific formula is as follows:

[0027] h k =f(h) k-1 ,x k (4)

[0028] Where: h k-1 Indicates the hidden state of the GRU neuron at time k-1; x k This represents the input to the GRU neuron at time k;

[0029] 5) The output of the Decoder architecture at time k-1 will be used as the input of the GRU neuron at time k, as shown in the following formula:

[0030] x k =U Gk-1 (5)

[0031] In the formula: U Gk-1 This represents the Decoder architecture output at time k-1;

[0032] 6) Substitute equation (5) into equation (4). Simultaneously, the Decoder architecture performs the opposite operation to the Encoder, decoding the intermediate state C of the input sequence step by step according to time steps to form the final output sequence, where the intermediate state C of the Decoder architecture at time k-1 is... k-1 with h k-1 Equal, at time k, the Decoder architecture output is determined by h. k-1 U Gk-1 and h k The decision was made jointly, and the details are as follows:

[0033]

[0034] In the formula: U Gk The output of the decoder architecture at time k is represented by p; the probability is represented by g; the softmax function is represented by F; and the transformation function is represented by F.

[0035] 7) Input x to the GRU neuron at time k k and the intermediate state C of the Decoder architecture at time k-1 k-1 For constructing the update gate z in the GRU neuron for the variable k Reset door r k and the output value to be determined The specific models for the three are as follows:

[0036]

[0037] In the formula: W r x represents k and r k Weighting coefficients between them; W z x represents k and z k Weighting coefficients between them; W h x represents k and The weighting coefficients between them; α represents the sigmoid activation function in the neural network;

[0038] 8) z k r k as well as Combining these three factors, we can obtain the hidden layer output h of the GRU neuron. k The specific formula is as follows:

[0039]

[0040] Where: h k-1 This represents the hidden layer output of the GRU neuron at time k-1;

[0041] The above steps are used to construct an Encoder-Decoder composite neural network architecture.

[0042] In step 3, the daily load data P of a typical day is... L and its corresponding unit start-up and shutdown scheme U G As a historical mapping sample, in a historical mapping sample, the unit start-up and shutdown scheme U G With daily load P L The relationship is represented by U. G =F(p(P) LThe description is as follows: p represents the probability between daily load and corresponding unit start-up and shutdown schemes, and F represents the transformation function.

[0043] Accumulate historical data and train a deep learning model based on Seq2Seq and GRU offline to obtain a model that can describe U. G With P L A mapping model of the probability relationship between them.

[0044] The Adam algorithm is used to train deep learning models, specifically through the following steps:

[0045] 1) Construct a loss function based on Mean Absolute Deviation (MAE) using an Encoder-Decoder architecture, and set the output of the Encoder-Decoder architecture at time k to U. Gok The target value is U Gdk The total error E of the sample during training is shown in the following formula:

[0046]

[0047] 2) The Adam algorithm is used as the neuron weight update algorithm to train the parameters of GRU neurons in the Encoder-Decoder architecture. Its basic formula is shown below.

[0048]

[0049] In the formula: θ k Let δ be the parameter variable to be updated at time k, and δ be the learning rate. and The weighted average value and biased variance of the gradient after error correction are calculated using the following formulas:

[0050]

[0051] 3) Substituting formula (11) into formula (10), and using the Adam algorithm to adaptively find the learning rate of each parameter, the learning rate of W in the GRU neurons within the Encoder-Decoder architecture can be adjusted. r W z and W h The adjustments to the three weighting coefficients are as follows:

[0052]

[0053] The Encoder-Decoder architecture is trained by continuously adjusting the weight coefficients using formula (12).

[0054] A composite neural network architecture for unit combination decision-making, based on GRU and Seq2Seq technology to construct an Encoder-Decoder composite neural network architecture, specifically includes the following steps:

[0055] 1) Transfer a historical mapping sample (P) L U G Substituting this into the Encoder-Decoder architecture, the Encoder architecture will convert the daily load sequence P... L The data is read in steps, and the hidden layer state of the GRU neuron at time t is jointly determined by the hidden layer state of the GRU neuron at time t-1 and the daily load at time t. The specific formula is as follows:

[0056] h t =f(h) t-1 ,P Lt (1)

[0057] Where: h t h represents the hidden state of the GRU neuron at time t; t-1 P represents the hidden state of the GRU neuron at time t-1; Lt This represents the daily load input at time t;

[0058] 2) In the Encoder architecture, the hidden state h of the GRU neuron at time t. t Similar to the intermediate states in the Encoder architecture, in the Decoder architecture, the hidden state h of the GRU neuron at time k is... k Similar to the intermediate states of the Decoder architecture, the specific formula is:

[0059]

[0060] In the formula: C t Represents the intermediate state of the Encoder architecture at time t; C k This represents the intermediate state of the Encoder architecture at time k;

[0061] 3) The intermediate state output by the Encoder architecture at time T is the intermediate state C of the input sequence, with a value of C. T , representing the complete information of the input sequence, specifically as follows:

[0062] C = C T (3);

[0063] 4) Input the intermediate state C of the sequence into the Decoder architecture. The initial value C0 of the Decoder intermediate state is the same as the intermediate state C of the sequence. After inputting C0, the hidden state h of the GRU neuron at time k can be obtained. kIt is jointly determined by the hidden state of the GRU neuron at time k-1 and the input of the GRU neuron at time k, and the specific formula is as follows:

[0064] h k =f(h) k-1 ,x k (4)

[0065] Where: h k-1 Indicates the hidden state of the GRU neuron at time k-1; x k This represents the input to the GRU neuron at time k;

[0066] 5) The output of the Decoder architecture at time k-1 will be used as the input of the GRU neuron at time k, as shown in the following formula:

[0067] x k =U Gk-1 (5)

[0068] In the formula: U Gk-1 This represents the Decoder architecture output at time k-1;

[0069] 6) Substitute equation (5) into equation (4). Simultaneously, the Decoder architecture performs the opposite operation to the Encoder, decoding the intermediate state C of the input sequence step by step according to time steps to form the final output sequence, where the intermediate state C of the Decoder architecture at time k-1 is... k-1 with h k-1 Equal, at time k, the Decoder architecture output is determined by h. k-1 U Gk-1 and h k The decision was made jointly, and the details are as follows:

[0070]

[0071] In the formula: U Gk The output of the decoder architecture at time k is represented by p; the probability is represented by g; the softmax function is represented by F; and the transformation function is represented by F.

[0072] 7) Input x to the GRU neuron at time k k and the intermediate state C of the Decoder architecture at time k-1 k-1 For constructing the update gate z in the GRU neuron for the variable k Reset door r k and the output value to be determined The specific models for the three are as follows:

[0073]

[0074] In the formula: W r x representsk and r k Weighting coefficients between them; W z x represents k and z k Weighting coefficients between them; W h x represents k and The weighting coefficients between them; α represents the sigmoid activation function in the neural network;

[0075] 8) z k r k as well as Combining these three factors, we can obtain the hidden layer output h of the GRU neuron. k The specific formula is as follows:

[0076]

[0077] Where: h k-1 This represents the hidden layer output of the GRU neuron at time k-1;

[0078] The above steps are used to construct an Encoder-Decoder composite neural network architecture.

[0079] A method for training a deep learning model of a power system, using the Adam algorithm, specifically employs the following steps:

[0080] 1) Construct a loss function based on Mean Absolute Deviation (MAE) using an Encoder-Decoder architecture, and set the output of the Encoder-Decoder architecture at time k to U. Gok The target value is U Gdk The total error E of the sample during training is shown in the following formula:

[0081]

[0082] 2) The Adam algorithm is used as the neuron weight update algorithm to train the parameters of GRU neurons in the Encoder-Decoder architecture. Its basic formula is shown below.

[0083]

[0084] In the formula: θ k Let δ be the parameter variable to be updated at time k, and δ be the learning rate. and The weighted average value and biased variance of the gradient after error correction are calculated using the following formulas:

[0085]

[0086] 3) Substituting formula (11) into formula (10), and using the Adam algorithm to adaptively find the learning rate of each parameter, the learning rate of W in the GRU neurons within the Encoder-Decoder architecture can be adjusted. r W z and W h The adjustments to the three weighting coefficients are as follows:

[0087]

[0088] The Encoder-Decoder architecture is trained by continuously adjusting the weight coefficients using formula (12).

[0089] Compared with existing methods, the data-driven unit combination decision-making method proposed in this invention has the following advantages and

[0090] Beneficial effects:

[0091] 1) This invention constructs a deep learning model for unit coordinating decision-making based on GRU, which has higher training efficiency compared with the LSTM model used in existing literature;

[0092] 2) Based on GRU, this invention introduces Seq2Seq technology and proposes an Encoder-Decoder composite neural network architecture for unit combination decision-making. Compared with the existing literature's method of using sample clustering, the method proposed in this invention does not require preprocessing of sample data for clustering. It can directly use a deep learning model to complete the training of all differential sample data, thus having higher training and decision-making efficiency.

[0093] 3) This invention proposes a sample encoding technique for high-dimensional unit combination sample matrices, which effectively reduces the dimensionality of unit combination sample data and further improves the training efficiency of unit combination deep learning models. Attached Figure Description

[0094] Figure 1 It is a data-driven unit combination decision-making framework based on a composite neural network architecture.

[0095] Figure 2 This is a schematic diagram of sample encoding technology.

[0096] Figure 3 The daily load and unit start-up and shutdown scheme mapping model.

[0097] Figure 4 It is an Encoder-Decoder composite neural network architecture.

[0098] Figure 5 This is a diagram of the internal structure of a GRU neuron.

[0099] Figure 6 This is a training error curve diagram based on the GRU model and based on Seq2Seq technology and the GRU model. Detailed Implementation

[0100] like Figure 1 As shown, the data-driven intelligent decision-making method for unit combination based on GRU and Seq2Seq technologies is characterized by the following steps:

[0101] 1. Use sample encoding techniques for high-dimensional unit combination training sample matrices to compress the dimensionality of historical decision-making data for unit combination;

[0102] 2. Based on threshold recurrent networks, Seq2Seq technology is introduced to establish a composite neural network architecture for unit combination decision-making;

[0103] 3. Based on this, a deep learning model for unit combination is constructed, and a mapping model between system daily load and unit start-up and shutdown schemes is established through training with historical data;

[0104] 4. Use the generated mapping model to make unit combination decisions, obtain the unit start-up and shutdown status and unit output under the optimal power flow model, and use the obtained unit combination decision results as new historical sample data to train the deep learning model, thereby realizing continuous improvement of the model.

[0105] like Figure 2 As shown, the daily load data P of a typical day is... L and its corresponding unit start-up and shutdown scheme U G As a historical mapping sample. In a mapping sample, the unit start-up and shutdown scheme U G With daily load P L The relationship can be represented by U G =F(p(P) L The mapping relationship is as follows: Figure 1 As shown.

[0106] Figure 2 In this context, p represents the probability between daily load and the corresponding unit start-up / shutdown plan, and F represents the transition function. For daily load P... L In essence, the process of building a mapping model involves accumulating a large amount of historical data and training a deep learning model based on Seq2Seq and GRU offline to obtain a model that can describe U. G With P L A mapping model of the probability relationship between them.

[0107] like Figure 3As shown, the start-stop state vector of the unit combination for each time period is encoded, and vectors with identical start-stop states have the same encoding. The main purpose of sample encoding is to convert the start-stop state vector of the unit combination for each time period into a corresponding decimal code, thereby compressing the dimension of the sample matrix and ultimately improving the training efficiency of the deep learning model. The principle is as follows: Figure 2 As shown.

[0108] An Encoder-Decoder composite neural network architecture is constructed based on GRU and Seq2Seq technologies. The specific architecture is as follows: Figure 4 As shown.

[0109] The structure of GRU neurons is as follows: Figure 5 As shown.

[0110] In the specific construction process, a historical mapping sample (P) will be used. L U G Substituting this into the Encoder-Decoder architecture, the Encoder architecture will convert the daily load sequence P... L The data is read in steps, and the hidden layer state of the GRU neuron at time t is jointly determined by the hidden layer state of the GRU neuron at time t-1 and the daily load at time t. The specific formula is as follows:

[0111] h t =f(h) t-1 ,P Lt (1)

[0112] Where: h t h represents the hidden state of the GRU neuron at time t; t-1 P represents the hidden state of the GRU neuron at time t-1; Lt This represents the daily load input at time t.

[0113] Based on the characteristics of the GRU model, in the Encoder architecture, the hidden state h of the GRU neuron at time t... t The intermediate states are the same as in the Encoder architecture. In the Decoder architecture, the hidden state h of the GRU neuron at time k is... k Similar to the intermediate states of the Decoder architecture, the specific formula is:

[0114]

[0115] In the formula: C t Represents the intermediate state of the Encoder architecture at time t; C k This represents the intermediate state of the Encoder architecture at time k.

[0116] Based on the characteristics of the encoder architecture, the intermediate state output by the encoder architecture at time T is the intermediate state C of the input sequence, and its value is C. T , representing the complete information of the input sequence, specifically as follows:

[0117] C = C T (3)

[0118] The intermediate state C of the sequence is input into the Decoder architecture, where the initial value C0 of the Decoder intermediate state is the same as the intermediate state C of the sequence. After inputting C0, the hidden layer state h of the GRU neuron at time k can be obtained. k It is jointly determined by the hidden state of the GRU neuron at time k-1 and the input of the GRU neuron at time k, and the specific formula is as follows:

[0119] h k =f(h) k-1 ,x k (4)

[0120] Where: h k-1 Indicates the hidden state of the GRU neuron at time k-1; x k This represents the input to the GRU neuron at time k.

[0121] Based on the characteristics of the Decoder architecture, the output of the Decoder architecture at time k-1 will be used as the input of the GRU neuron at time k, as shown in the following formula:

[0122] x k =U Gk-1 (5)

[0123] In the formula: U Gk-1 This represents the Decoder architecture output at time k-1.

[0124] Substituting equation (5) into equation (4), the Decoder architecture performs the opposite operation to the Encoder, decoding the intermediate state C of the input sequence step by step according to time steps to form the final output sequence, where the intermediate state C of the Decoder architecture at time k-1 is... k-1 with h k-1 They are equal. Therefore, the Decoder architecture output at time k is determined by h. k-1 U Gk-1 and h k The decision was made jointly, and the details are as follows:

[0125]

[0126] In the formula: U Gk denoted by k; p represents the output of the Decoder architecture at time k; g represents the softmax function; and F represents the transformation function.

[0127] Depend on Figure 4 It can be seen that, with x as the input to the GRU neuron at time k... k and the intermediate state C of the Decoder architecture at time k-1 k-1 For constructing the update gate z in the GRU neuron for the variable k Reset door r k and the output value to be determined The specific models for the three are as follows:

[0128]

[0129] In the formula: W r x represents k and r k Weighting coefficients between them; W z x represents k and z k Weighting coefficients between them; W h x represents k and The weighting coefficients between them; α represents the sigmoid activation function in the neural network.

[0130] z k r k as well as Combining these three factors, we can obtain the hidden layer output h of the GRU neuron. k The specific formula is as follows:

[0131]

[0132] Where: h k-1 This represents the output of the hidden layer of the GRU neuron at time k-1.

[0133] In step 4, the Adam algorithm is used to train the deep learning model.

[0134] The Encoder-Decoder architecture uses the Mean Absolute Deviation (MAE) as the basis for its loss function, and sets the output of the Encoder-Decoder architecture at time k to U. Gok The target value is U Gdk The total error E of the sample during training is shown in the following formula.

[0135]

[0136] The Adam algorithm is used as the neuron weight update algorithm to train the parameters of GRU neurons in the Encoder-Decoder architecture. Its basic formula is shown below.

[0137]

[0138] In the formula: θ k Let δ be the parameter variable to be updated at time k, and δ be the learning rate. and The weighted average value and biased variance of the gradient after error correction are calculated using the following formulas.

[0139]

[0140] Substituting equation (11) into equation (10), and using the Adam algorithm to adaptively find the learning rate for each parameter, the learning rate of W in the GRU neurons within the Encoder-Decoder architecture can be adjusted. r W z and W h The adjustments to the three weighting coefficients are as follows.

[0141]

[0142] The Encoder-Decoder architecture is trained by continuously adjusting the weight coefficients using formula (12).

[0143] Example:

[0144] To verify the effectiveness and correctness of this invention, simulation tests were conducted based on the IEEE 118-node standard example and actual data from the Hunan power grid. In the IEEE 118-node example, based on the daily load characteristic curve of the Hunan power grid, load samples for 93 typical days applicable to the IEEE 118-node system were constructed. Of these daily load samples, samples 1-90 were used as training samples, and samples 91-93 were used as test samples. For ease of subsequent calculation and analysis, samples 1-90 can be clustered into three sample sets using the method described in the literature "Research on Data-Driven Intelligent Decision-Making Method for Unit Combination with Self-Learning Capability": samples 1-30 belong to cluster set 1, and test sample 91 also belongs to this type; samples 31-60 belong to cluster set 2, and test sample 92 also belongs to this type; samples 61-90 belong to cluster set 3, and test sample 93 also belongs to this type.

[0145] All deep learning models for unit combinations in this invention were trained and tested on the Tensorflow 1.6.0 platform. All related simulation calculations were performed on a computer with an Intel Core i5-4460 processor / 3.20GHz and 8GB of memory.

[0146] To verify the correctness of the method of the present invention, four methods were set up: Method 1 is a unit combination decision method based on LSTM model; Method 2 is a unit combination decision method based on GRU model; Method 3 is a unit combination decision method based on Seq2Seq technology and GRU model; Method 4 is a unit combination decision method based on Seq2Seq technology and GRU model that incorporates sample coding technology.

[0147] 1) Process simulation and correctness verification of the method of the present invention

[0148] First, the method of this invention was used to train samples 1-90. The trained mapping model was then used to make unit combination decisions for test samples 91-93. The resulting unit start-up and shutdown schemes were compared with the unit start-up and shutdown schemes based on the physical model-driven unit combination decision method in the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach". The solution results for test sample 91 are shown in Table 1.

[0149] Table 1. Unit start-up and shutdown schemes obtained from solving test sample 91 using the methods of this invention and the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach".

[0150]

[0151] As shown in Table 1, the unit start-up and shutdown schemes obtained by the method of this invention and the method in the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach" are the same, indicating that this invention can fully learn the mapping relationship between daily load and unit start-up and shutdown schemes. The trained mapping model can make correct unit start-up and shutdown scheme decisions for any type of input daily load data.

[0152] For test samples 91-93, this invention, based on the obtained start-up and shutdown schemes, solves the optimal power flow model to obtain the unit combination decision results. The unit combination decision results of this invention's method are compared with those of the method in the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach," and the total cost comparison is shown in Table 2.

[0153] Table 2. Comparison of total costs between the method of this invention and the method in the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach".

[0154]

[0155] Table 2 shows that the unit output scheme and total cost obtained by the method of this invention are the same as those obtained by the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach". These results indicate that after obtaining the unit start-up and shutdown schemes, the method of this invention can obtain the same unit output scheme as the method in the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach" by solving the optimal power flow model.

[0156] This is due to the unit start-up and shutdown scheme U. G In both the method of this invention and the method in the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach", these are decision variables. When solving for unit output schemes using both the method of this invention and the method in the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach", the same optimal power flow model is used. Therefore, for the same unit start-up and shutdown scheme U... GThe method of this invention and the method in the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach" can solve for the same unit output scheme P. G The self-learning and self-evolution capabilities of the data-driven intelligent unit combination decision-making method, as well as its applicability to different types of unit combination problems, have been verified in the literature "Research on Data-Driven Intelligent Unit Combination Decision-Making Method with Self-Learning Capability," and will not be elaborated upon here. The decision accuracy shown in subsequent examples represents the correlation between the unit start-up and shutdown schemes obtained by the method of this invention and those obtained in the literature "Network-Constrained AC Unit Commitment Under Uncertainty: A Benders' Decomposition Approach."

[0157] 2) Validation of the effectiveness of the GRU model introduced in this invention.

[0158] To verify the effectiveness of the GRU model introduced in this invention, Method 1 and Method 2 were first trained using the training samples after cluster preprocessing, with the number of training iterations set to 500. Then, the above two methods were used to solve the test samples 91-93. The specific results are shown in Table 3.

[0159] Table 3 Comparison of decision accuracy and training time between Method 1 and Method 2

[0160]

[0161] As shown in Table 3, in terms of decision accuracy, Method 2 achieved 100% accuracy when making decisions on all three test samples, while Method 1's accuracy was below 100% when making decisions on sample number 93. Furthermore, Method 1's total cost was higher than Method 2's. This indicates that after 500 training iterations, Method 2 was able to generate accurate mapping models for all cluster training sample sets, while Method 1 could not yet generate accurate mapping models for cluster sample set 3, requiring more training iterations. In terms of training time, compared to Method 1, Method 2 reduced the training time for the three cluster sample sets by 77s, 91s, and 82s, respectively. This demonstrates that, with the same number of training iterations, Method 2 requires less training time.

[0162] The main reason for the above phenomenon is that GRU merges the input gate and the forget gate, reorganizes the forget gate into an update gate and a reset gate, and simplifies the memory units in LSTM, enabling it to directly calculate and output results. Therefore, the overall structure of the model is simpler, and its training and decision-making accuracy is higher under the same training parameters. Thus, it is clear that this invention's use of GRU to replace LSTM in constructing a deep learning model for unit combination decision-making is correct and effective.

[0163] 3) Verification of the effectiveness of the Seq2Seq technology introduced in the method of this invention

[0164] To verify the effectiveness of the Seq2Seq technique introduced in the method of this invention, samples that have undergone clustering preprocessing and those that have not been clustered preprocessing were used as training samples for methods 2 and 3, respectively. The two methods were then used to solve for test samples 91-93, and the specific results are shown in Table 4.

[0165] Table 4 Comparison of decision accuracy and training time between Method 2 and Method 3

[0166]

[0167] As shown in Table 4, regarding decision-making accuracy, if the training samples of the unit combination are not pre-processed by clustering and are directly used to train Method 2, the decision-making accuracy of Method 2 after training is generally below 90%. Training Method 2 with training samples that have undergone clustering preprocessing achieves 100% decision-making accuracy. Method 3, however, is different; even when trained with unprocessed training samples, it still achieves 100% decision-making accuracy. This indicates that after introducing Seq2Seq technology, a single deep learning model can complete the training of all dissimilar sample data. The reason is that if a single deep learning model is directly constructed using GRU, a unique compromise mapping model will inevitably be generated when faced with highly dissimilar training sample data, making it difficult to guarantee the accuracy of online decision-making. Therefore, to ensure online decision-making accuracy, the training samples must first undergo clustering preprocessing, and then a corresponding deep learning model is used to train each class of training samples. If we introduce Seq2Seq technology to construct a GRU-based Encoder-Decoder composite neural network architecture, since its intermediate state C can completely preserve the category information and pointing probability of the input and output sequences, we can use a single deep learning model to achieve accurate training on all differential samples.

[0168] In terms of training time, training Method 2 using unclustered samples requires the least total time. While training Method 3 using unclustered samples takes 110.02 seconds longer than Method 2 using clustered samples, it still takes 179.23 seconds less. This is because training Method 2 using unclustered samples will prematurely terminate the training process as it fails to converge to the most accurate mapping model, thus requiring less overall training time. However, training Method 2 using clustered samples requires training multiple deep learning models, and the clustering preprocessing also takes time, resulting in the longest total training time in this case.

[0169] To further analyze the reasons for the difference in training time between Method 2 and Method 3, the actual convergence curves of the two methods are given when training with unclustered samples, as shown below. Figure 6 As shown.

[0170] Depend on Figure 6 It can be seen that the total training error of Method 2 converges to around 0.09 after more than 100 training iterations, and the error cannot be reduced further, thus the training process ends prematurely. In contrast, the error of Method 3 eventually converges to around 0.0002 after more than 100 training iterations. Therefore, it is evident that because it requires more training iterations, Method 3 takes longer to train on unclustered samples, while Method 2, although having a shorter training time, cannot guarantee the training accuracy of the model.

[0171] In summary, directly training Method 2 with training samples that have not undergone clustering preprocessing makes it difficult to guarantee the model's decision-making accuracy. While introducing clustering preprocessing can address the training accuracy issue of Method 2 when dealing with discrepancies in training samples, it significantly increases the complexity of offline training, thus greatly increasing both the required offline training time and the total training time. Method 3, by incorporating Seq2Seq technology, allows for accurate training on discrepancies using only a single deep learning model. The training process is simpler, and it improves the training and decision-making efficiency of the deep learning model while maintaining training accuracy.

[0172] 4) Verification of the effectiveness of the sample coding technique introduced in this invention

[0173] To verify the impact of the sample encoding technique proposed in this invention on the training efficiency of deep learning models, methods 3 and 4 were trained using unclustered training samples, and the training and testing results are shown in Table 5.

[0174] Table 5 Comparison of decision accuracy and training time between Method 3 and Method 4

[0175]

[0176] As shown in Table 5, after training, Method 3 and Method 4 achieve the same decision accuracy, both reaching 100% accuracy when making decisions on three test samples. However, compared to Method 3, Method 4 requires 351 seconds less training time. This is because the sample encoding technique proposed in this invention can directly compress the data dimension of the training samples, reducing the unit start-stop state matrix of one training sample from 24×54 dimensions to 24×1 dimensions. This directly reduces the number of variables that the deep learning model needs to calculate during training, thereby reducing the model's training time. Therefore, although the sample encoding process consumes some time after introducing the sample encoding technique, this invention can effectively reduce the overall training time of the deep learning model.

[0177] In summary, the sample encoding technique proposed in this invention can effectively compress the data dimensionality of unit combination training samples, directly reducing the number of variables that the deep learning model needs to calculate during training, thereby effectively reducing its training time while ensuring the training accuracy of the deep learning model. Compared with the LSTM neural network used in the literature "Research on Intelligent Decision-Making Method for Unit Combination with Data-Driven Self-Learning Capability", the GRU model introduced in this invention can achieve higher training and decision-making accuracy under the same training parameters. By introducing Seq2Seq technology, this invention constructs an Encoder-Decoder composite neural network architecture with GRU as neurons, which enables the method of this invention to achieve accurate training of differential samples with only a single deep learning model. The training process is simpler, and it can improve the training and decision-making efficiency of the method of this invention while ensuring its training accuracy.

Claims

1. A method for training a deep learning model of a power system, characterized in that, The Adam algorithm is used to train deep learning models, specifically through the following steps: Step 1: Construct a loss function based on Mean Absolute Error (MAE) using an Encoder-Decoder architecture, and set... k The output of the Encoder-Decoder architecture at any given time is U Gok The target value is U Gdk Then the total error of the sample during training E As shown in the following formula: （9）； Step 2: Use the Adam algorithm as the neuron weight update algorithm to train the parameters of the GRU neurons in the Encoder-Decoder architecture. The basic formula is shown below. （10）； In the formula: θ k for k Parameter variables that need to be updated at all times δ For learning rate, and The weighted average value and biased variance of the gradient after error correction are calculated using the following formulas: （11）； Step 3: Substitute formula (11) into formula (10), and use the Adam algorithm to adaptively find the learning rate of each parameter to achieve the learning rate of GRU neurons in the Encoder-Decoder architecture. W r , W z as well as W h The adjustments to the three weighting coefficients are as follows: （12）； The Encoder-Decoder architecture is trained by continuously adjusting the weight coefficients using formula (12); A mapping model between system daily load and unit start-up and shutdown schemes is established through training; the mapping model describes the unit start-up and shutdown schemes. U G Daily load P L Probabilistic relationships between them; unit start-up and shutdown schemes U G Daily load P L Relationship U G = F ( p ( P L ))describe; in, p This represents the probability between daily load and corresponding unit start-up / shutdown plans. F Indicates the transformation function; The deep learning model is derived from a composite neural network architecture for unit combination decision-making, which is an Encoder-Decoder composite neural network architecture built on GRU and Seq2Seq technologies. The composite neural network architecture for unit combination decision-making includes the following steps during construction: 1) Transfer a historical mapping sample ( P L , U G Substituting this into the Encoder-Decoder architecture, the Encoder architecture will convert the daily load sequence... P L Step-by-step reading, in t The hidden state of the GRU neuron at time t is determined by t The hidden layer state of GRU neurons at time -1 and t The time of day and daily load are jointly determined, and the specific formula is as follows: （1）； In the formula: h t express t The hidden state of a GRU neuron at any given time; h t-1 express t- The hidden layer state of the GRU neuron at time 1; P Lt express t Daily load input at any time; 2) In the Encoder architecture, t Time-bound GRU neuron hidden state h t Similar to the intermediate states in the Encoder architecture, in the Decoder architecture, the intermediate states are... k Time-bound GRU neuron hidden state h k Similar to the intermediate states of the Decoder architecture, the specific formula is: （2）； In the formula: C t express t Intermediate state of the encoder architecture at any given moment; C k express k Intermediate state of the encoder architecture at any given moment; 3) Encoder architecture in T The intermediate state of the output at any given time is the intermediate state of the input sequence. C The value is C T , representing the complete information of the input sequence, specifically as follows: （3）； 4) Input the intermediate state C of the sequence into the Decoder architecture, where the initial value of the Decoder intermediate state C0 is the same as the intermediate state of the sequence. C The same, will C After inputting 0, the result is... k Time-bound GRU neuron hidden state h k , its origin k- Time 1 GRU neuron hidden layer state and k The time interval is jointly determined by the inputs to the GRU neurons, and the specific formula is as follows: （4）； In the formula: h k-1 express k- The hidden layer state of the GRU neuron at time 1; x k express k Time-dependent input to GRU neurons; 5) k The Decoder architecture output at time -1 will be used as... k The input to the GRU neuron at time t is as follows: （5）； In the formula: U Gk-1 express k Decoder architecture output at time -1; 6) Substitute equation (5) into equation (4). At the same time, the Decoder architecture performs the opposite operation to the Encoder, converting the intermediate states of the input sequence into... C Decoding is performed step-by-step according to time steps to form the final output sequence, where k -1 is the intermediate state of the Decoder architecture. C k-1 and h k-1 equal, k The output of the Decoder architecture at any moment is determined by h k-1 , U Gk-1 as well as h k The decision was made jointly, and the details are as follows: （6）； In the formula: U Gk express k Output of the Decoder architecture at any moment; p Represents probability; g This represents the softmax function; F Indicates the transformation function; 7) with k Time-dependent GRU neuron input x k and k- 1-Time Decoder Architecture Intermediate State C k-1 Constructing update gates in GRU neurons for variables z k Reset door r k and the output value to be determined The specific models for the three are as follows: （7）； In the formula: W r express x k and r k Weighting coefficients between them; W z express x k and z k Weighting coefficients between them; W h express x k and Weighting coefficients between them; This represents the sigmoid activation function in a neural network. 8) z k , r k as well as The combined results of these three factors yield the hidden layer output of the GRU neuron. h k The specific formula is as follows: （8）； In the formula: h k-1 express k- Output of the hidden layer of the GRU neuron at time 1.

Images