A power grid load sequence prediction method and system based on a generative adversarial network

Abstract

The application provides a power load sequence prediction method and system based on a generative adversarial network, which encodes meteorological temperature and humidity and economic data and power load data influencing a power grid load through an LSTM, and enables the generative adversarial network to learn the coupling effect between the related factors in a pure data-driven manner, and finally outputs a future sequence prediction value by using an LSTM decoder. The application can accurately perform power grid load prediction, and solves the problems of single power grid load prediction sequence, short time step, low precision and the like.

Description

Technical Field

[0001] This invention relates to a method and system for predicting power load sequences based on generative adversarial networks, belonging to the field of power grid load sequence prediction technology. Background Technology

[0002] Reliable and accurate load forecasting facilitates dynamic planning for smart buildings and smart grids, enabling targeted demand response strategies and investments in power infrastructure. Improving the accuracy of load forecasting has been a key research focus for many years. Because load characteristics are easily influenced by multiple factors, and their changes do not exhibit clear patterns, traditional load characteristic analysis methods only select the impact of one factor on load characteristics, focusing on a single mapping analysis between a single factor and load characteristics. This neglects the relationship between electricity load and major influencing factors, resulting in inaccurate load characteristic analysis and impacting load forecasting and electricity consumption planning.

[0003] Deep learning methods can better handle the volatility and randomness of load data, improving the accuracy of power grid load forecasting to some extent. However, current load forecasting applications often screen forecast samples before building the model, using similarity to select training samples. This leads to the loss of information contained in unselected samples. The accuracy of the forecast results is also affected by the selected samples. When the selected samples are inaccurate, the forecast accuracy decreases, and it cannot meet the load forecasting requirements for load data with rapidly changing distribution. Summary of the Invention

[0004] The purpose of this invention is to provide a power load sequence prediction method and system based on generative adversarial networks, which solves the problems of single power grid load prediction sequence, short time step and low accuracy.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] This invention provides a power load sequence prediction method based on generative adversarial networks, comprising:

[0007] The real sequence of meteorological temperature and humidity, economic data and power load data before the current moment is input into the constructed generative adversarial network to obtain the predicted value of the future sequence.

[0008] The best predicted sequence is obtained by filtering the predicted values ​​of the future sequence.

[0009] Furthermore, the process of inputting the real sequence of meteorological temperature and humidity, economic data, and electricity load data prior to the current moment into the constructed generative adversarial network to obtain predicted values ​​for future sequences includes:

[0010] A fixed-length vector is obtained by embedding sample data consisting of meteorological temperature and humidity, economic data, and power load data prior to the current moment using a single-layer sensor.

[0011] The resulting vector After performing LSTM encoding, we get:

[0012]

[0013] The interaction between meteorological temperature and humidity, economic data, and electricity load data is learned to obtain the set tensor P of the sample data. i ;

[0014] right and P i Decoding yields predicted values ​​for the future sequence:

[0015]

[0016]

[0017]

[0018]

[0019] Where φ(.) is an embedding function with nonlinearity. Let represent the meteorological temperature and humidity, economic data, and electricity load data of the i-th sample at time t, respectively. Let γ(.) be the predicted value of the future sequence, and W be a multilayer perceptron with ReLU nonlinearity. ee and W ed All are embedding weights, W encoder represents the LSTM weights, and N is the number of samples.

[0020] Furthermore, it also includes:

[0021] The discriminator of the generative adversarial network is fed into the real sequence composed of predicted values ​​of future sequences and sample data. The weight parameters of the generative adversarial network are updated and adjusted based on the discrimination results.

[0022] Furthermore, the process of filtering the predicted values ​​of the future sequence to obtain the optimal predicted sequence includes:

[0023] The optimal prediction sequence is obtained based on the following diversity loss function:

[0024]

[0025] Among them, Y i Represents the true sequence. For the prediction sequence, k represents the k output prediction values ​​obtained by random sampling.

[0026] This invention also provides a power load sequence prediction system based on generative adversarial networks, including a generative adversarial network and an output module.

[0027] The generative adversarial network includes a generator and a discriminator;

[0028] The generator includes an encoder, a coordination layer, and a decoder;

[0029] The encoder is used to encode meteorological temperature and humidity, economic data, and power load data based on the LSTM unit;

[0030] The overall planning layer is used to learn the interaction between meteorological temperature and humidity, economic data and power load data to obtain a set tensor of sample data;

[0031] The decoder is used to decode LSTM encoded data and set tensors, and output the predicted value of the future sequence;

[0032] The discriminator is used to make a judgment based on the predicted value of the future sequence output by the generator and the real sequence composed of sample data, and to optimize and adjust the weight parameters of the generative adversarial network.

[0033] The output module is used to filter the predicted values ​​of future sequences to obtain the best predicted sequence.

[0034] Furthermore, the encoder is specifically used for,

[0035] A fixed-length vector is obtained by embedding sample data consisting of meteorological temperature and humidity, economic data, and power load data prior to the current moment using a single-layer sensor.

[0036] The resulting vector Inputting into the LSTM unit yields:

[0037]

[0038] Where φ(.) is an embedding function with nonlinearity. W represents the meteorological temperature and humidity, economic data, and electricity load data of the i-th sample at time t. ee For embedding weights, W encoder These are the weights for the LSTM.

[0039] Furthermore, the decoder is specifically used for,

[0040] right Decoding the set tensor is as follows:

[0041]

[0042]

[0043]

[0044]

[0045] in, Let γ(.) be the predicted value of the future sequence, and let P be a multilayer perceptron with ReLU nonlinearity. i Let W be the set tensor of the i-th sample. ed The embedding weights are N, where N is the number of samples.

[0046] Furthermore, the output module is specifically used for,

[0047] The optimal prediction sequence is obtained based on the following diversity loss function:

[0048]

[0049] Among them, Y i Represents the true sequence. For the prediction sequence, k represents the k output prediction values ​​obtained by random sampling.

[0050] The beneficial effects of this invention are as follows:

[0051] This invention provides a power load sequence prediction method based on generative adversarial networks (GANs). It employs LSTM encoding for meteorological temperature and humidity, economic data, and power load, all factors influencing power grid load. Furthermore, it proposes a novel interaction model within the latent space of the GAN, enabling the network to accurately represent the mapping relationship between these factors and power load, thus achieving more accurate power grid load prediction. This method addresses the problems of single power grid load prediction sequences, short time steps, and low accuracy. Attached Figure Description

[0052] Figure 1 This is a diagram of the generative adversarial network structure proposed in this invention.

[0053] Figure 2 A flowchart of a power load sequence prediction method based on generative adversarial networks is provided for an embodiment of the present invention. Detailed Implementation

[0054] The present invention will now be further described. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0055] This invention provides a power load sequence prediction method based on generative adversarial networks (GANs). By modeling the interactions between relevant factors affecting load and jointly analyzing the sequences of all prediction units in the prediction sample, a generative adversarial network GAN-based encoder-decoder architecture is proposed. Meteorological temperature and humidity, economic data, and power load, which affect grid load, are encoded using LSTM, and a novel interaction model is proposed in the latent space. Finally, the grid load prediction result is obtained using an LSTM decoder and used as input to the GAN network discriminator. By training the network in a purely data-driven manner, the network learns the coupling effects between relevant factors, enabling it to accurately represent the mapping relationship between meteorological temperature and humidity, economic data, and power load, thus achieving accurate grid load prediction.

[0056] A power load sequence prediction method based on generative adversarial networks includes:

[0057] The real sequence of meteorological temperature and humidity, economic data and power load data before the current moment is input into the constructed generative adversarial network to obtain the predicted value of the future sequence.

[0058] The best predicted sequence is obtained by filtering the predicted values ​​of the future sequence.

[0059] A preferred embodiment of the present invention provides a power load sequence prediction method based on generative adversarial networks, see [link to relevant documentation]. Figure 2 The details are as follows:

[0060] Step 1: Define variables. Factors affecting power grid load include meteorological temperature and humidity Te and economic data Ec. These two types of factors affecting power grid load, along with the power load data Lo, are selected as network variables X = [Te, Ec, Lo].

[0061] The known information of the sample is processed from time step t = 1, ..., t obs Initially, define N samples X = X1, X2, ..., X... N ,

[0062] The subscript i represents the i-th sample, and the superscript t represents time.

[0063] From the time step t = t obs +1,...,t pred The future sequence is defined as:

[0064] use This represents the predicted value of the future sequence.

[0065] Step two: Construct a Generative Adversarial Network (GAN). A GAN consists of two adversarially trained neural networks. These two adversarially trained models are: a generative model G to capture the data distribution, and a discriminative model D to estimate the probability that a sample comes from the training data rather than G. The generative model G takes a latent variable z as input and outputs a sample G(z). The discriminative model D takes a sample x as input and outputs D(x) representing its true probability. The training objective function is as follows:

[0066]

[0067] GANs can generate G(z,c) and D(x,c) by providing a generator and a discriminator with additional perturbation factors as input c.

[0068] Step 3: Design the model framework. The generative model can be used with time series data to simulate the probability of future sequences.

[0069] See Figure 1 The model mainly consists of two parts: a generator and a discriminator. The most crucial component in the generator is the coordination module. Based on an encoding / decoding framework, the generator connects the hidden states of the encoder and decoder through the coordination module.

[0070] The discriminator input includes the input sequence X. i and future predictions (or Y) i The entire sequence of ) is analyzed and classified as "true / false".

[0071] Specifically as follows:

[0072] Generator. First, the load values ​​are embedded using a single-layer MLP to obtain a fixed-length vector. These embeddings are used as LSTM units input to the encoder at time t, introducing the following recursion:

[0073]

[0074]

[0075] Where φ(.) is an embedding function with nonlinearity, W ee It is the embedding weight, W encoder These are the LSTM weights shared among all prediction units in the sample.

[0076] The overall planning layer. Each sample unit (a set of meteorological temperature, humidity, and economic data constitutes one sample unit) cannot capture the interactions between relevant factors using a single LSTM. The encoder understands the state of each region and memorizes their historical sequences. To synthesize information from different encoders to effectively predict correlations, this invention uses an overall planning module to model the interactions between relevant factors.

[0077] In t obs Then the state of all samples at the scene was hidden, and t obs The previous sample set yielded the set tensor P for each sample. i To achieve simultaneous prediction across multiple samples, an information-sharing mechanism is needed among LSTMs. However, the following problems need to be addressed: (1) Due to the large number of varying prediction units globally, a concise representation that integrates information from all samples is required. (2) The prediction unit information is numerous and complex; integrating information from only one location is insufficient, therefore, modeling the global configuration is necessary. The first problem was solved by proposing a grid-based coordination scheme, but this manual solution is slow and cannot capture the global environment. The above properties can be achieved by applying a symmetric function trained on the transformation elements of the input point set.

[0078] Decoder. Traditional GANs generate samples using noise as input. However, the goal of this invention is to generate future scenarios consistent with the past. To achieve this goal, the generation conditions of the generator's output sequence are set by initializing the hidden state of the decoder as follows:

[0079]

[0080]

[0081] In the formula, For the output sequence, γ(.) is a multilayer perceptron (MLP) with ReLU nonlinearity, W c It is the embedding weight, and z is the potential factor that affects the sample output result. In reality, there will be some perturbation factors.

[0082] After initializing the decoder state as described above, a fixed-length vector is obtained. The following sequence predictions can be obtained:

[0083]

[0084]

[0085]

[0086]

[0087] Where φ(.) is an embedding function with ReLU nonlinearity, W ed As an embedding weight, W decoder This represents the LSTM weights.

[0088] The discriminator consists of a single encoder.

[0089] Specifically, the discriminator will T real =[X i ,Y i ]or As input, these are classified as true / false. In this invention, MLP is applied to the last hidden state of the encoder for classification. Ideally, the discriminator learns the rules of interaction between the predictor units and classifies sequences that do not meet the interaction rules as "false". In addition to adversarial loss, this invention applies L2 loss to the predicted sequences of the generated samples and the real data.

[0090] Step 4: Sequence prediction.

[0091] Within a finite past history, the model must consider multiple possible outcomes. The methods described so far have produced good predictions, but these predictions attempt to generate an "average" prediction given the possibility of multiple outputs. Furthermore, we have found that the outputs are not very sensitive to variations in noise and produce very similar predictions. Therefore, this invention proposes a diversity loss function to encourage the network to generate diverse samples. For each sample, k possible output predictions are generated by randomly sampling from N(0,1), and the "best" prediction in the L2 sense is selected as the final prediction:

[0092]

[0093] Here, k is a hyperparameter.

[0094] This loss encourages the network to hedge its choices by considering only the best sequence and covers the output space loss that conforms to past sequences. Structurally, it is similar to the Minimum (MoN) loss over N. This invention reports two error metrics: Average Displacement Error (ADE) and Final Displacement Error (FDE). This invention consistently outperforms state-of-the-art S-LSTM methods, especially for long-term predictions (lower is better). LSTMs are used as RNNs in both the decoder and encoder models. The hidden state dimensions of the encoder and decoder are 16 and 32, respectively. Input coordinates are embedded as 16-dimensional vectors. The generator and discriminator are trained iteratively using Adam with an initial learning rate of 0.001 and a batch size of 64, suitable for 200 iterations.

[0095] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0096] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0098] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting power load sequences based on generative adversarial networks, characterized in that, include: The actual sequence of meteorological temperature and humidity, economic data, and electricity load data prior to the current moment is input into a constructed generative adversarial network to obtain predicted values ​​for future sequences, including: A fixed-length vector is obtained by embedding sample data consisting of meteorological temperature and humidity, economic data, and power load data prior to the current moment using a single-layer sensor. The resulting vector After performing LSTM encoding, we get: The interaction between meteorological temperature and humidity, economic data, and electricity load data is learned to obtain the set tensor P of the sample data. i ; right and P i Decoding yields predicted values ​​for the future sequence: Where φ(.) is an embedding function with nonlinearity. Let represent the meteorological temperature and humidity, economic data, and electricity load data of the i-th sample at time t, respectively. Let γ(.) be the predicted value of the future sequence, and W be a multilayer perceptron with ReLU nonlinearity. ee and W ed All are embedding weights, W encoder represents the LSTM weights, and N is the number of samples; The best predicted sequence is obtained by filtering the predicted values ​​of future sequences, including: The optimal prediction sequence is obtained based on the following diversity loss function: Among them, Y i Represents the true sequence. For the prediction sequence, k represents the k output prediction values ​​obtained by random sampling.

2. The power load sequence prediction method based on generative adversarial networks according to claim 1, characterized in that, Also includes: The discriminator of the generative adversarial network is fed into the real sequence composed of predicted values ​​of future sequences and sample data. The weight parameters of the generative adversarial network are updated and adjusted based on the discrimination results.

3. A power load sequence prediction system based on generative adversarial networks, characterized in that, To implement the power load sequence prediction method based on generative adversarial networks as described in claim 1, the system includes a generative adversarial network and an output module. The generative adversarial network includes a generator and a discriminator; The generator includes an encoder, a coordination layer, and a decoder; The encoder is used to encode meteorological temperature and humidity, economic data, and power load data based on the LSTM unit; The overall planning layer is used to learn the interaction between meteorological temperature and humidity, economic data and power load data to obtain a set tensor of sample data; The decoder is used to decode LSTM encoded data and set tensors, and output the predicted value of the future sequence; The discriminator is used to make a judgment based on the predicted value of the future sequence output by the generator and the real sequence composed of sample data, and to optimize and adjust the weight parameters of the generative adversarial network. The output module is used to filter the predicted values ​​of future sequences to obtain the best predicted sequence.

4. The power load sequence prediction system based on generative adversarial networks according to claim 3, characterized in that, The encoder is specifically used for, A fixed-length vector is obtained by embedding sample data consisting of meteorological temperature and humidity, economic data, and power load data prior to the current moment using a single-layer sensor. The resulting vector Inputting into the LSTM unit yields: Where φ(.) is an embedding function with nonlinearity. W represents the meteorological temperature and humidity, economic data, and electricity load data of the i-th sample at time t. ee For embedding weights, W encoder These are the weights for the LSTM.

5. A power load sequence prediction system based on generative adversarial networks according to claim 4, characterized in that, The decoder is specifically used for, right Decoding the set tensor is as follows: in, Let γ(.) be the predicted value of the future sequence, and let P be a multilayer perceptron with ReLU nonlinearity. i Let W be the set tensor of the i-th sample. ed The embedding weights are N, where N is the number of samples.

6. The power load sequence prediction system based on generative adversarial networks according to claim 3, characterized in that, The output module is specifically used for, The optimal prediction sequence is obtained based on the following diversity loss function: Among them, Y i Represents the true sequence. For the prediction sequence, k represents the k output prediction values ​​obtained by random sampling.

Images