Method and device for short-term power load forecasting during infectious disease epidemic

Abstract

The application discloses a kind of infectious disease epidemic period short-term power load prediction method and device, by constructing including different geographical area power load data, time information, weather data and mobile data knowledge base, with the knowledge base trained migration deep learning model, the input of the migration deep learning model is the time information, weather data and mobile data of different geographical area, the output of the migration deep learning model is the short-term power load prediction result of different geographical area, in training with depth reinforcement learning to determine the optimal hyperparameter of the migration deep learning model shared weight layer, with the migration deep learning model of optimal hyperparameter trained to carry out power load prediction.The application can improve prediction accuracy, and overcome the problem of limited mobile data related to infectious disease epidemic.

Description

Technical Field

[0001] This application belongs to the field of power load forecasting technology, and in particular relates to a method and apparatus for short-term power load forecasting during an infectious disease epidemic. Background Technology

[0002] Short-term load forecasting (STLF) refers to predicting electricity demand from one hour to one week. The predicted short-term load demand helps improve the efficiency of power system dispatch. Although accurate short-term load forecasting is challenging due to the uncertainty and volatility of electricity demand, some deep learning models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have achieved good prediction accuracy.

[0003] Due to the susceptibility to infectious diseases, many countries and regions have implemented measures requiring strict social distancing restrictions. These restrictions can lead to a significant drop in electricity demand in a short period, making it difficult for traditional electricity load forecasting models to accurately predict electricity demand during infectious disease outbreaks. Furthermore, traditional deep learning models applied to short-term electricity load forecasting typically use past electricity load demand values, time-series information, and weather data as input features. However, deep learning models using these traditional features struggle to capture changes in electricity load demand during infectious disease outbreaks because they do not incorporate relevant social and economic information. Summary of the Invention

[0004] The purpose of this application is to provide a method and apparatus for short-term power load forecasting during infectious disease outbreaks. A knowledge base containing mobile data is constructed, reflecting changes in passenger flow in different areas and buildings based on mobile services. The socioeconomic behavioral information contained in the mobile data enables the forecasting model to predict significant drops in power load over a short period.

[0005] To achieve the above objectives, the technical solution of this application is as follows:

[0006] A method for short-term power load forecasting during an infectious disease outbreak includes:

[0007] Construct a knowledge base that includes power load data, time information, weather data, and mobility data from different geographical regions;

[0008] A transfer deep learning model is trained using a constructed knowledge base. The inputs to the transfer deep learning model are time information, weather data, and mobility data from different geographical regions. The output of the transfer deep learning model is the short-term power load forecast results for different geographical regions. During training, deep reinforcement learning is used to determine the optimal hyperparameters of the shared weight layer of the transfer deep learning model. The state of the deep reinforcement learning is the hyperparameters of the shared weight layer of the transfer deep learning model. The action of the deep reinforcement learning is the specific action of adjusting the hyperparameters. The reward of the deep reinforcement learning is the prediction accuracy correlation function.

[0009] Power load forecasting is performed using a trained transfer deep learning model with optimal hyperparameters.

[0010] Furthermore, the hyperparameters include the number of iterations, batch size, and learning rate.

[0011] Furthermore, the prediction accuracy is the mean absolute percentage error.

[0012] Furthermore, the step of determining the optimal hyperparameters of the shared weight layer of the transfer deep learning model during training using deep reinforcement learning also includes:

[0013] The extreme gradient boosting machine is trained by using the states traversed by deep reinforcement learning and their corresponding Q-values ​​as the training set.

[0014] The Q-values ​​of states not traversed by deep reinforcement learning are predicted using the trained extreme gradient boosting machine, and the state corresponding to the maximum Q-value is selected as the optimal value of the hyperparameters.

[0015] Furthermore, the weight sharing layer employs an RNN recurrent neural network layer.

[0016] This application also proposes a short-term power load forecasting device during an infectious disease outbreak, including a processor and a memory storing a number of computer instructions, which, when executed by the processor, implement the steps of the short-term power load forecasting method during an infectious disease outbreak.

[0017] This application proposes a method and apparatus for short-term power load forecasting during infectious disease outbreaks. Real datasets covering 12 different countries and cities were used to verify the performance of the proposed model.

[0018] Based on the results of multiple comparative experiments, the following four conclusions can be drawn:

[0019] (1) Information on changes in socioeconomic behavior contained in the knowledge base helps deep learning models predict sudden drops in power load during infectious disease outbreaks. This also provides a new direction for predicting power load under other global emergencies.

[0020] (2) By making full use of information on changes in socioeconomic behavior in the knowledge base, the transfer deep learning model proposed in this application can overcome the problem of limited mobile data related to infectious disease outbreaks.

[0021] (3) The proposed hyperparameter optimization method based on reinforcement learning can automatically optimize the hyperparameters of transfer deep learning models.

[0022] (4) The proposed advance prediction method improves the efficiency of hyperparameter optimization of reinforcement learning agents by predicting the state-behavior values ​​of untraversed state spaces. Attached Figure Description

[0023] Figure 1 This is a flowchart of the short-term power load forecasting method during an infectious disease outbreak as described in this application;

[0024] Figure 2 This is a schematic diagram illustrating the advance prediction of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] In one embodiment, such as Figure 1 As shown, a method for short-term power load forecasting during an infectious disease outbreak is provided, including:

[0027] Step S1: Construct a knowledge base that includes power load data, time information, weather data, and mobility data for different geographical regions.

[0028] The knowledge base constructed in this application comprises four normalized data categories: electricity load data, time information, weather data, and mobility data. The data in this knowledge base covers 12 different geographical regions, including the United Kingdom (UK), Germany, France, the California Independent System Operator (CAISO) region, the NYISO region, Dallas, Houston, San Antonio (SA), Boston, Chicago, Philadelphia, and Seattle. The data spans from February 15, 2020 to May 15, 2020, covering the period before and after the implementation of strict social distancing restrictions.

[0029] (1) The power load data represents the hourly power load demand values ​​for different geographical regions. Among them, the power load data for Europe comes from European transmission network operators, and the power load data for the United States comes from the respective independent system operators in the United States.

[0030] (2) Time information refers to the day of the week and time information represented by unique hot coding.

[0031] (3) Weather data comes from World Weather Online and includes information such as cloud cover, humidity, precipitation, air pressure, and temperature.

[0032] (4) Mobile data from Google and Apple revealed relative changes in visitor traffic across different areas and buildings. Google's mobile data revealed relative changes in visitor numbers at six different locations: retail and entertainment venues, grocery stores and pharmacies, parks, transit hubs, workplaces, and residential areas. The baseline for the data was the median for the five weeks from January 3, 2020 to February 6, 2020. Mobile data from Apple revealed relative changes in visitor numbers for three types of movement: driving, commuting, and walking. The baseline for the data was the relevant data as of January 13, 2020. Both Google and Apple collect information based on the historical location of user accounts.

[0033] Step S2: Train a transfer deep learning model using the constructed knowledge base. The input of the transfer deep learning model consists of time information, weather data, and mobility data from different geographical regions. The output of the transfer deep learning model is the short-term power load forecast results for different geographical regions. During training, deep reinforcement learning is used to determine the optimal hyperparameters of the shared weight layer of the transfer deep learning model. The state of the deep reinforcement learning is the hyperparameters of the shared weight layer of the transfer deep learning model. The action of the deep reinforcement learning is the specific action of adjusting the hyperparameters. The reward of the deep reinforcement learning is the prediction accuracy correlation function.

[0034] This application employs a constructed knowledge base to train a transfer deep learning model, and then uses the trained model to predict short-term power load in a target area. The transfer deep learning model addresses the problem of insufficient mobile data related to the pandemic. This model can utilize transfer knowledge learned from the source domain to predict other target learning tasks and can solve the problem of limited mobile data during infectious disease outbreaks. In this application, data from different source domains refers to temporal information, weather data, and mobile data from different geographical regions, and different learning tasks refer to the problem of short-term power load prediction in different geographical regions.

[0035] The definition of transfer deep learning is as follows: Given a source domain D S Source learning task T S Target domain DT and target learning task T T In D S ≠D T T S ≠T T In this case, transfer deep learning models utilize data from D... S and T S Transfer knowledge learned in middle school to improve the target prediction function r T (·) in D T The learning ability within. According to the above definition, each domain of transfer deep learning is defined as D = {F, P(X)}, where F = {f1, ..., f...} n} is an n-dimensional feature space, X = {x1, ..., x2} n Let F ∈ F be the learning samples, and P(X) be the marginal probability distribution of X. The feature space and marginal probability distribution are different in different domains. Each learning task is defined as T={y,r(·)}, where y is the value space of the actual electricity load demand, and r(·) is the prediction function.

[0036] In one specific embodiment, the transfer deep learning model includes an input layer, a shared weight layer, a hidden layer, and an output layer. This application optimizes the hyperparameters of the shared weight layer, including the number of iterations, batch size, and learning rate, to improve the prediction accuracy of the transfer deep learning model. Furthermore, transfer learning addresses the problem of limited mobile data during infectious disease outbreaks. In one specific embodiment, the weight sharing layer uses an RNN (Recurrent Neural Network) layer, and the hidden layer uses a fully connected layer. RNNs are recurrent neural networks, which perform well in processing continuous temporal features; moreover, RNNs have fewer hyperparameters, resulting in lower time complexity during reinforcement learning parameter tuning. Other neural networks, such as LSTM, CNN, and ResNet, can also be used for the weight sharing layer.

[0037] This application uses data from different source domains (a knowledge base containing mobile data from 12 regions) to train the above-mentioned transfer deep learning model and obtains the network weights of the final model.

[0038] Finally, the data for different learning tasks (electricity load prediction tasks for 12 regions) are sequentially input into the trained transfer deep learning model. The data from different regions are processed in the hidden layer after passing through the weight sharing layer, and finally the prediction results for different learning tasks are obtained.

[0039] In order to obtain the optimal hyperparameters of the shared weight layer of the transfer deep learning model, this application uses deep reinforcement learning (Q-learning) to optimize the hyperparameters of the shared weight layer.

[0040] Deep reinforcement learning models problems as Markov decision processes, in which a reinforcement learning agent learns through trial and error in the environment. The goal of the reinforcement learning agent is to select an action that maximizes the expected discounted reward. This application uses deep reinforcement learning methods to optimize the hyperparameters of the proposed transfer learning model to improve its prediction accuracy. A Markov decision process is defined as...<S,A,P,R> S refers to the set of all possible valid states of the reinforcement learning agent, where each state consists of different numbers of iterations, batch sizes, and learning rates. s refers to a specific state, where... A refers to the set of all possible actions the reinforcement learning agent can perform. At each time point, the reinforcement learning agent will execute one of six candidate actions, and accordingly change the values ​​of the number of iterations, batch size, or learning rate. 'a' refers to a specific action, where... P refers to the transition probability distribution of the reinforcement learning agent in the constructed environment. R refers to the reward function of the reinforcement learning agent. t represents different points in time, k represents the future time step, and r t+k γ represents the reward obtained by the reinforcement learning agent after k time steps. γ is a discount factor that balances the importance of immediate and future rewards. This application uses the prediction accuracy on the test set as an example. t As a signal for updating the reward of a reinforcement learning agent, r t =1 - accuracy t , and r t ∈(0,1).

[0041] Q π (s,a) is defined as a state-behavior value, representing the expected cumulative discount reward obtained by performing action a according to policy π:S→A in state s, as shown in formula (1):

[0042]

[0043] Among them, s t and a t Let π and t represent the state and action of the reinforcement learning agent at time t, respectively. This application employs the Q-learning algorithm to continuously estimate the optimal state-action values ​​in order to obtain the optimal policy π. * (s)∈argmax a Q * (s,a), Q * (s,a) refers to the optimal state-behavior value. The state-behavior value can also be represented by the Q-value. The core formula in the Q-learning algorithm, namely the Bellman equation, is shown in formula (2):

[0044]

[0045] The pseudocode for the hyperparameter optimization method based on reinforcement learning is shown in Algorithm 1, where Q(·,·) represents the set of state-behavior values ​​at all time points, and Q t+1 (s t ,a t The value of is represented as the Q value at time t+1. In reinforcement learning, a training cycle refers to a complete round trip of the reinforcement learning agent's interaction with the environment, which includes different time steps.

[0046]

[0047] Table 1

[0048] In Q-learning, the values ​​of the number of iterations, batch size, and learning rate are initialized as the state s at time t. t At each time point, action a is obtained using the ε-greedy algorithm. t a t It is an integer whose different values ​​represent different operations for the size of the hyperparameter value, as shown in Table 2.

[0049]

[0050]

[0051] Table 2

[0052] With state s t Run the transfer deep learning model, where the hyperparameters are the number of iterations, batch size, and learning rate at time t, and obtain the prediction accuracy of the transfer deep learning model at time t, which is expressed by the mean absolute percentage error (MAPE), as shown in formula (3):

[0053]

[0054] Where N is the total number of hours predicted, y i and p i , respectively, represent the actual and predicted power load values ​​for the i-th hour.

[0055] The reward r at time t is obtained based on the prediction accuracy of the transfer deep learning model at time t. t and perform the selected action a t , and obtain the new state s t+1This involves setting the new set of iteration count, batch size, and learning rate. Then, the Q-value and the state at time t+1 are updated, and the above steps are repeated until the pre-defined termination time step is reached. The state (iteration count, batch size, and learning rate) with the largest Q-value among all time points is selected and represented as the optimized hyperparameter value.

[0056] It should be noted that the specific implementations of transfer learning and Q-learning are relatively mature technologies in this field, and will not be elaborated here.

[0057] Step S3: Use the trained transfer deep learning model with optimal hyperparameters to predict power load.

[0058] After obtaining the optimal hyperparameters, a transfer learning model is trained using these optimal hyperparameters. This trained model can then be used to predict short-term power load. During prediction, time information, weather data, and mobility data are input, and the predicted power load is output.

[0059] In one specific embodiment, to improve the efficiency of hyperparameter optimization for reinforcement learning agents and thus obtain better hyperparameters, a pre-prediction method is employed in deep reinforcement learning, including:

[0060] The extreme gradient boosting machine is trained by using the states traversed by deep reinforcement learning and their corresponding Q-values ​​as the training set.

[0061] The Q-values ​​of states not traversed by deep reinforcement learning are predicted using the trained extreme gradient boosting machine, and the state corresponding to the maximum Q-value is selected as the optimal value of the hyperparameters.

[0062] In this embodiment, it is implemented using extreme gradient boosting (XGBoost). Figure 2 As shown, each Q-value represents the state-behavior value of the reinforcement learning agent at different time points, whether it has traversed or not.

[0063] In this embodiment, the transfer deep learning model using a hyperparameter optimization method based on reinforcement learning traverses a subset of states and obtains the corresponding Q-values. Figure 2 The gray squares in the diagram represent states visited by the reinforcement learning agent and their corresponding Q-values. The white squares represent states not visited by the agent and their corresponding Q-values. The states visited by the agent and their corresponding Q-values ​​are divided into training and test sets for training XGBoost. Finally, the trained XGBoost model is used to predict the Q-values ​​of states not visited by the agent. Figure 2The hyperparameters are represented by squares with diagonal lines. Then, the state corresponding to the largest Q value is selected, which is represented as the optimal value of the hyperparameters.

[0064] This application also verifies the prediction accuracy of the model through experiments. The prediction results of the RNN model (RNN_KB) using the knowledge base constructed in this application and the RNN model (RNN) without using the knowledge base are shown in Table 3:

[0065]

[0066] Table 3

[0067] As shown in Table 3, the prediction accuracy of the RNN model using a knowledge base is higher than that of the RNN model without a knowledge base. The prediction accuracy improvements for France, Germany, the UK, NYISO, CAISO, Dallas, Houston, San Antonio, Boston, Chicago, Philadelphia, and Seattle are 40.5%, 49.2%, 12.5%, 64.9%, 57.5%, 9.1%, 33.8%, 0.2%, 58.5%, 70.6%, 75.1%, and 56.3%, respectively. Experiments demonstrate that knowledge bases, especially those containing mobile data, can effectively improve the accuracy of deep learning models in predicting electricity load during infectious disease outbreaks.

[0068] The prediction results for the transfer learning deep learning model and the RNN model without transfer learning (RNN_KB) are shown in Table 4. It should be noted that the prediction results for RNN_KB differ from those in Table 3 because the results in Table 4 were obtained by rerunning the model five times.

[0069]

[0070] Table 4

[0071] As shown in Table 4, compared to RNN models without transfer learning, transfer deep learning models achieved higher prediction accuracy on eight datasets (France, Germany, UK, Dallas, Houston, Boston, Philadelphia, and Seattle), but lower accuracy on only four datasets (NYISO, CAISO, San Antonio, and Chicago). Overall, transfer deep learning models outperform RNN models without transfer learning, which can be attributed to their ability to leverage socioeconomic behavioral changes contained within mobile data.

[0072] In the hyperparameter optimization based on reinforcement learning and using an advance prediction method, the values ​​of the initial parameters in the experiment ranged as follows: number of iterations [0, 30], batch size [30, 60], and learning rate [0, 0.005]. The prediction results for different training periods and time steps are shown in Table 5. Different combinations of training periods and time steps indicate the number of states the reinforcement learning agent can traverse. For example, a combination of 5 training periods and 10 time steps means the reinforcement learning agent can traverse a maximum of 50 states. The variables ep, bs, and lr represent the optimal values ​​of the number of iterations, batch size, and learning rate for different combinations of training periods and time steps, respectively. State_num represents the number of states traversed by the reinforcement learning agent. Time represents the time cost required to train the proposed model under different training periods and time steps. MAPE is used to evaluate the prediction accuracy of the model.

[0073]

[0074]

[0075] Table 5

[0076] The prediction results in Tables 4 and 5 show that the prediction accuracy of this application is higher than that of ordinary transfer deep learning models, achieving improvements of 14.6%, 13.6%, 26.2%, 24.6%, 21.6%, 64.5%, 53.4%, 71.0%, 20.1%, 17.7%, 30.1%, and 37.2% respectively on the 12 datasets: France, Germany, UK, NYISO, CAISO, Dallas, Houston, San Antonio, Boston, Chicago, Philadelphia, and Seattle. The use of the pre-prediction method significantly improves the hyperparameter optimization efficiency of the reinforcement learning agent. For example, on the French dataset, with a training epoch of 30 and a time step of 30, the reinforcement learning agent can traverse up to 900 (30×30) states. However, the reinforcement learning agent actually only traversed 608 states and obtained the corresponding Q-values, taking a total of 39,765 seconds (i.e., 11.05 hours) to traverse the 608 states. The Q-values ​​of the remaining 292 (900-608) states were predicted using the advance prediction method. Since these 292 states were not actually traversed, it can be considered that using the advance prediction method saved approximately 19097 seconds (292 / 608*39765) of time cost (i.e. 5.30 hours).

[0077] In one embodiment, this application also provides a short-term power load forecasting device during an infectious disease outbreak, including a processor and a memory storing a plurality of computer instructions, which, when executed by the processor, implement the steps of the short-term power load forecasting method during an infectious disease outbreak.

[0078] Specific limitations regarding short-term power load forecasting devices during infectious disease outbreaks can be found in the above section on the limitations of short-term power load forecasting methods during infectious disease outbreaks, and will not be repeated here. The aforementioned short-term power load forecasting devices during infectious disease outbreaks can be implemented entirely or partially through software, hardware, or a combination thereof. They can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0079] The memory and processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The memory stores a computer program that can run on the processor, which implements the network topology layout method in this embodiment of the invention by running the computer program stored in the memory.

[0080] The memory may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory stores the program, and the processor executes the program upon receiving an execution instruction.

[0081] The processor may be an integrated circuit chip with data processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0082] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for short-term electric power load forecasting during an infectious disease epidemic, characterized by, The short-term power load prediction method during the infectious disease epidemic period comprises the following steps: A knowledge base including power load data of different geographical regions, time information, weather data and mobile data, which is relative change data of passenger flow, is constructed; A transfer deep learning model is trained by using the constructed knowledge base, input of the transfer deep learning model is time information, weather data and mobile data of different geographical regions, output of the transfer deep learning model is short-term power load prediction results of different geographical regions, in the training, deep reinforcement learning is used to determine optimal hyperparameters of a shared weight layer of the transfer deep learning model, state of the deep reinforcement learning is the hyperparameters of the shared weight layer of the transfer deep learning model, action of the deep reinforcement learning is a specific action of adjusting the hyperparameters, reward of the deep reinforcement learning is a prediction accuracy related function; Power load prediction is performed by using the trained transfer deep learning model with optimal hyperparameters; In the training, the deep reinforcement learning is used to determine the optimal hyperparameters of the shared weight layer of the transfer deep learning model, further comprising: States and corresponding Q values traversed by the deep reinforcement learning are used as a training set to train an extreme gradient boosting machine; Q values of states not traversed by the deep reinforcement learning are predicted by using the trained extreme gradient boosting machine, a state corresponding to a maximum Q value is selected as an optimal value of the hyperparameters.

2. The short-term power load forecasting method during an infectious disease outbreak according to claim 1, characterized in that, The hyperparameters include iteration number, batch size and learning rate.

3. The method of claim 1, wherein the method is characterized by, The prediction accuracy is mean absolute percentage error.

4. The short-term power load forecasting method during an infectious disease outbreak according to claim 1, characterized by, The shared weight layer adopts an RNN recurrent neural network layer.

5. An apparatus for short-term electric power load forecasting during an infectious disease epidemic, comprising a processor and a memory having stored therein a plurality of computer instructions, wherein, The computer instructions are executed by the processor to realize steps of the method in any one of claims 1 to 4.

Images