A method, device and storage medium for industry-specific power load forecasting based on federated learning

Abstract

The present invention discloses a method, device and storage medium for power load forecasting by industry based on federated learning. The method comprises the steps of (1) analyzing factors affecting power load; (2) preprocessing original power load data; (3) constructing an industry user data set according to the load influencing factors; (4) establishing a load forecasting model based on LSTNet; and (5) establishing a load forecasting model by industry based on federated learning based on the FedAvg algorithm and the FedML framework.

Description

Technical Field

[0001] The present invention relates to power technology, and in particular to a method, device and storage medium for forecasting power load by industry based on federated learning. Background Art

[0002] my country's energy is accelerating its transformation towards green and low-carbon development. A large number of distributed power generation and energy storage devices are connected to the distribution network, and the traditional distribution network is gradually transforming into an active distribution network. However, with the gradual opening of the electricity sales market under the active distribution network, a large amount of user-side data has been collected and used by all parties due to its huge application value in load forecasting and other aspects. Existing research can no longer meet users' urgent needs for privacy and security. Therefore, research on the privacy protection of electricity users needs to be promoted urgently.

[0003] Traditional load forecasting is based on the research of time series forecasting. Time series forecasting uses historical time series to predict future time series. It mainly includes methods such as ARIMA, VAR, random forest (RF), gated recurrent unit (GRU), extreme learning machine (ELM), and long short-term memory (LSTM). These methods require the use of electricity consumption data collected by devices such as smart meters to perform load forecasting, which is very likely to expose user privacy. Summary of the Invention

[0004] This invention aims to provide a method, device, and storage medium for industry-specific power load forecasting based on federated learning. First, the factors influencing load are analyzed, and weather, economic factors, and time of day are selected as load correlation factors. Based on this analysis, the raw power load data is preprocessed to construct an industry user dataset. A load forecasting model is established based on LSTNet. Furthermore, a federated learning-based industry-specific load forecasting model is established using the FedAvg algorithm and the FedML framework.

[0005] The purpose of the present invention can be achieved by the following technical solution: A method for forecasting power load by industry based on federated learning, comprising the following steps:

[0006] Step 1: Construct an industry user dataset based on load influencing factors;

[0007] Step 2: Establish a load forecasting model based on LSTNet;

[0008] Step 3: Establish an industry-specific load forecasting model based on federated learning based on the FedAvg algorithm and the FedML framework.

[0009] Optionally, the load influencing factors include: weather factors, economic factors and time factors.

[0010] Optionally, the raw power load data is preprocessed before step 1, comprising the following steps:

[0011] (1) Use the 3-Sigma criterion to detect possible outliers in the load data set;

[0012] (2) Fill in missing values according to formulas (1)-(2):

[0013] x ij =X i-1 (1-w ij )+X i w ij ,i=2,3,...,T,j=1,2,...,N (1)

[0014]

[0015] Among them, X is the data value of the original sequence at hour i, x ij is the jth data value in the i-th hour of the interpolated sequence, w ij For X i →x ij The weight of , N is the total number of split points within 1 hour, and T is the total number of hours in the original sequence;

[0016] (3) Use mean variance normalization to scale features:

[0017]

[0018] Among them, x i , μ, σ represent the i-th input data corresponding to a certain feature, the mean of the input data, and the variance of the input data respectively;

[0019] (4) Set an input sequence sliding window of length P and an output sequence sliding window of length N, and apply these two sliding windows to the time series to obtain an input sequence x of length P. t-1 ,...,x t-P and an output sequence x of length N t ,...,x t+N-1 ;

[0020] After the input sequence is input into the time series prediction model, the predicted sequence is obtained, as shown in formula (4):

[0021]

[0022] Among them, x t represents the input value at time t, xt represents the predicted value at time t, x t-1 ,...,x t-P represents an input sequence of length P, Represents a prediction sequence of length N. The difference between the output sequence and the prediction sequence is quantified by the evaluation index, which can evaluate the prediction performance of the time series prediction model.

[0023] Optionally, step 1 of constructing an industry user data set based on load influencing factors includes the following steps:

[0024] Load, weather and time are selected as the characteristic types of input data for the load forecasting model; weather data includes temperature, relative humidity, precipitation, wind speed and air pressure; time data includes holiday information, weekday information, year, month, day, hour, minute and day of the week information.

[0025] Optionally, the step 2 of establishing a load forecasting model based on LSTNet includes the following steps:

[0026] (1) Convolutional layer setting d c The convolution kernel of size (m,n) is used, where m is the width of the convolution kernel and n is the number of features. The kth convolution kernel scans the input sequence and outputs h of length T. k Vector, as shown in formula (5):

[0027] h k =RELU(W k *X+b k ) (5)

[0028] Among them, * is the convolution operation, W k and b k are weight coefficients and bias respectively. The output matrix of the convolutional layer is composed of d c h k Vector composition, size (d c ,T), the corresponding weight coefficients and bias are W C and b C ;

[0029] (2) The recurrent layer is a GRU model. The unit structure of GRU consists of an update gate and a reset gate. The activation function is RELU. The formulas are shown in (6)-(9):

[0030] r t =σ(W r [h t-1 ,x t ]+b r ) (6)

[0031] z t =σ(W z[h t-1 ,x t ]+b z ) (7)

[0032] c t =RELU(W c [r t *h t-1 ,x t ]+b c ) (8)

[0033] h t =(1-z t )*h t-1 +z t *c t (9)

[0034] Among them, r t 、z t and c t are the reset gate coefficient, update gate coefficient and candidate hidden layer coefficient at time t respectively; W r 、W z and W c are the weight coefficients of the reset gate, update gate and candidate hidden layer respectively; b r 、b z and b c are the reset gate, update gate and bias of the candidate hidden layer respectively; h t-1 and h t are the output values of the recurrent layer at time t-1 and time t respectively;

[0035] (3) The recurrent-jump layer is a GRU model, and p is introduced to capture the periodicity of the time series. The formulas are shown in Equations (10)-(13):

[0036] r t =σ(W r [h t-p ,x t ]+b r ) (10)

[0037] z t =σ(W z [h t-p ,x t ]+b z ) (11)

[0038] c t =RELU(W c [r t *h t-p ,x t ]+b c ) (12)

[0039] h t =(1-z t )*h t-p +z t *c t (13)

[0040] Where p is the number of skipped hidden units, h t-p is the output value of the loop-skip layer at time tp. In order to distinguish the loop layer from the loop-skip layer, let is the output value of the recurrent layer at time t, is the output value of the cyclic skip layer at time t;

[0041] (4) The fully connected layer receives the output of the recurrent layer and the recurrent-skip layer, as shown in formula (14):

[0042]

[0043] in, is the output value of the fully connected layer at time t, is the output value of the recurrent-jump layer at time ti, W R and W i S are the weight coefficients of the recurrent layer at time t and the recurrent-skip layer at time i, b D is the bias of the fully connected layer. ;

[0044] (5) The autoregressive layer is responsible for providing the linear component, which helps improve the prediction efficiency in large-scale data. As shown in formula (15):

[0045]

[0046] in, is the output of the autoregressive layer, q ar is the size of the input window acting on the input sequence, W ar and b ar are the weight coefficient and bias of the autoregressive layer respectively;

[0047] (6) The final output prediction result is the superposition of the output of the neural network and the linear component of the autoregressive layer, as shown in formula (16):

[0048]

[0049] Among them, y t is the prediction result at time t. The weight coefficient corresponding to the convolution layer is W C , bias is b C , the weight coefficient corresponding to the loop layer is W R , bias is b R, the weight coefficient corresponding to the cycle-jump layer is W S , bias is b S , the bias of the fully connected layer is b D , the weight coefficient corresponding to the autoregressive layer is W ar , bias is b ar . The weight coefficients of each layer in LSTNet are W={W C ,W R ,W S ,W ar} and bias b={b C ,b R ,b S ,b D ,b ar} is recorded as

[0050] ω∈{W,b} (17) where ω represents the weight coefficient and bias of neurons in each layer, i.e., the model parameters of LSTNet.

[0051] Optionally, step 3 establishes an industry-specific load forecasting model based on federated learning based on the FedAvg algorithm and the FedML framework, including the following steps:

[0052] (1) Let interaction round t = 0, 2, ..., T be the number of local model parameters transmitted by the responding user to the central server located at the active distribution network operator, local training batch e = 1, 2, ..., E be the number of local training iterations of the responding user based on the LSTNet load forecasting model, and model parameter ω be expressed as Equation (18),

[0053]

[0054] ω t is the weight coefficient or bias of neurons in each layer of LSTNet after the local training batch reaches E in the tth round of interaction.

[0055] (2) The central server located at the active distribution network operator sends the initialized model parameters ω0 to all responding users;

[0056] (3) After the user receives the model parameters, the LSTNet model is trained locally based on the gradient descent algorithm;

[0057] (4) In the tth round of interaction, respond to user k to update ω t k Then it is sent to the central server located at the active distribution network operator;

[0058] (5) The central server collects all the model parameters updated by the responding users After that, aggregate the parameters to generate global parameters

[0059]

[0060] Among them, n k is the number of samples of responding user k, and n is the total number of samples of all responding users.

[0061] (6) The central server sets the global parameter ω t Sent to all responding users, who then update their local parameters based on the global parameters

[0062]

[0063] Then continue training until the Tth interaction is completed.

[0064] A device comprising:

[0065] one or more processors;

[0066] a memory for storing one or more programs;

[0067] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned industry-specific power load forecasting method based on federated learning.

[0068] A storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform any of the above-mentioned federated learning-based industry-specific power load forecasting methods.

[0069] Beneficial effects: The present invention predicts the electricity load of power users on the basis of fully considering privacy protection, which enables users in the same industry to conduct federal training without sharing load data, effectively protecting users' electricity privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0071] Figure 1 Schematic diagram of the flow of the industry-specific power load forecasting method based on federated learning in an embodiment of the present invention;

[0072] Figure 2 This is the LSTNet structure diagram;

[0073] Figure 3 Provides an overall framework for industry-specific load forecasting based on federated learning;

[0074] Figure 4 The comparison results of the load forecast values of the traditional scheme and the proposed scheme for the pharmaceutical manufacturing industry over two days;

[0075] Figure 5 The comparison results of the two-day load forecast values of the traditional solution and the solution proposed in this paper for the food manufacturing industry.

[0076] Figure 6 It is a structural diagram of a device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0077] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0078] This embodiment provides a method for forecasting power load by industry based on federated learning. Figure 1 As shown, including:

[0079] Step 1: Analyze the factors affecting power load;

[0080] Step 2: Preprocessing the original power load data;

[0081] Step 3: Construct an industry user data set based on load influencing factors;

[0082] Step 4: Establish a load forecasting model based on LSTNet;

[0083] Step 5: Establish an industry-specific load forecasting model based on federated learning based on the FedAvg algorithm and the FedML framework.

[0084] The following is a specific example to illustrate the execution process of the industry-specific power load forecasting method based on federated learning provided by an embodiment of the present invention and the beneficial effects that can be achieved.

[0085] The industry user load dataset used contains the electricity consumption data of industry users in various cities in Jiangsu Province from January 1, 2015 to December 31, 2016. The data interval is 15 minutes, and there are 96 records per day.

[0086] In order to consider the impact of weather in load forecasting, actual local weather data for the same period is also required. The weather data of a certain area from January 1, 2015 to December 31, 2016 is selected. Due to the frequency of data collection, the data interval is 1 hour, and there are 24 records per day, covering temperature (℃), relative humidity (%), precipitation (mm), wind speed (m / s 2 ) and air pressure (Pa).

[0087] After comprehensively considering the factors influencing load, load data, weather data, and time data were ultimately selected as input information for the load forecasting model. Weather data, including temperature, relative humidity, precipitation, wind speed, and air pressure, reflects the impact of weather factors on load. Time data, including holiday information, weekday information, and year, month, day, hour, minute, and day of the week, reflects the cyclical nature of load. Holiday information is based on my country's statutory holidays in 2015 and 2016; weekday information includes mid-week days and workdays resulting from adjusted holidays.

[0088] Furthermore, the original power load data is preprocessed by performing anomaly detection, missing value filling, normalization, sliding window processing, etc. to construct an industry user data set, as shown in Table 1:

[0089] Table 1 Feature selection and feature transformation of dataset

[0090]

[0091] The load forecasting model is established based on LSTNet, which consists of three parts: convolutional layer, recurrent layer and recurrent-skip layer, fully connected layer and autoregressive layer. Figure 2 shown.

[0092] Based on the FedAvg algorithm and FedML framework, a federated learning-based industry load forecasting model is established, such as Figure 3 The main bodies include the active distribution network operator, the central server set up by the active distribution network operator, the user, and the power sales company. The entire process is divided into 7 steps, as shown below:

[0093] (1) The active distribution network operator sends a training request to the users corresponding to a target industry;

[0094] (2) Considering that users need to complete training locally, users can decide whether to participate in federated training. Participating users are responding users. Responding users pre-process their own data sets locally and input the data sets into the LSTNet load forecasting model for local training. After completing a round of training, the neuron parameters of each layer of LSTNet are transmitted to the central server.

[0095] (3) The central server aggregates the local model parameters passed by the responding user based on the FedAvg algorithm to generate global model parameters and passes the global model parameters to the responding user. After multiple interactions, the central server obtains the industry global model;

[0096] (4) The central server transmits the industry global model to the active distribution network operator;

[0097] (5) The active distribution network operator distributes corresponding rewards based on the contribution of the responding users to the industry global model;

[0098] (6) The electricity sales company submits a model demand application to the active distribution network operator based on the required industries;

[0099] (7) Active distribution network operators return the required industry global model and obtain benefits.

[0100] The federated learning training scenario is set up according to Table 2, where the interaction round refers to the number of times the central server of each industry updates the global parameters to each responding user, and the local training batch refers to the number of local iterations of each responding user. The load forecasting model based on industry electricity consumption characteristics is trained on an industry-by-industry basis. Two industries are set up, namely the pharmaceutical manufacturing industry and the food manufacturing industry. Each industry uses data sets from three responding users. Based on the FedAvg algorithm, the data sets of different responding users in the same industry are integrated to capture the common electricity consumption characteristics of a single industry and obtain the final industry load forecasting model. The training method without considering industry electricity consumption characteristics is to have six responding users regardless of industry, and the local model parameters from different responding users are aggregated by the same central server, and finally a global model applicable to the six responding users is obtained.

[0101] Table 2 Federated learning training scenario settings

[0102]

[0103] RMSE and SMAPE are selected as the evaluation indicators of short-term load forecasting results, and their expressions are shown in Equations (21) and (22):

[0104]

[0105]

[0106] Where N represents the total number of samples, and y i Represent the predicted value and actual value at time i, respectively. RMSE uses the square term to amplify the gap between large and small errors, making it more sensitive to data with large forecast deviations. A smaller RMSE indicates better model prediction accuracy, and a smaller SMAPE indicates better forecast performance.

[0107] The prediction accuracy of federated learning in different scenarios is shown in Tables 3 and 4. The RMSE and SMAPE values are low, indicating high prediction accuracy.

[0108] Table 3 Comparison of RMSE prediction accuracy in different scenarios of federated learning

[0109]

[0110] Table 4 Comparison of prediction accuracy SMAPE in different scenarios of federated learning

[0111]

[0112] Figure 3 and Figure 4 The comparison of the predicted values for two days in the pharmaceutical and food manufacturing industries is shown in the figure below. The predicted values of the proposed solution are closer to the actual values than those of the traditional solution, which does not consider privacy protection and accumulates data first and then predicts.

[0113] Figure 6 This is a structural diagram of a device provided by an embodiment of the present invention, which provides services for implementing the industry-specific power load forecasting method based on federated learning in the above-mentioned embodiment of the present invention. Figure 6 A block diagram of an exemplary device 12 suitable for use in implementing embodiments of the present invention is shown. Figure 6 The device 12 shown is only an example and should not bring any limitation to the functionality and scope of use of the embodiments of the present invention.

[0114] like Figure 6 As shown, device 12 is implemented as a general-purpose computing device. Components of device 12 may include, but are not limited to, one or more processors or processing units 16, a system memory 28, and a bus 18 that connects various system components (including system memory 28 and processing unit 16).

[0115] Bus 18 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or a local bus using any of a variety of bus architectures. Examples of these architectures include, but are not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MAC) bus, an Enhanced ISA bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus.

[0116] Device 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by device 12, including volatile and non-volatile media, removable and non-removable media.

[0117] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. Device 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be configured to read and write to non-removable, non-volatile magnetic media. Each drive may be connected to bus 18 via one or more data media interfaces. Memory 28 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of various embodiments of the present invention.

[0118] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data, each of which, or some combination thereof, may include an implementation of a network environment. Program modules 42 generally implement the functions and / or methods of the embodiments described herein.

[0119] The device 12 may also communicate with one or more external devices 14 (e.g., a keyboard, a pointing device, a display 24, etc.), one or more devices that enable a user to interact with the device 12, and / or any device that enables the device 12 to communicate with one or more other computing devices (e.g., a network card, a modem, etc.). Such communication may occur through an input / output (I / O) interface 22. Furthermore, the device 12 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network such as the Internet) through a network adapter 20. Figure 6 As shown, network adapter 20 communicates with the other modules of device 12 via bus 18. It should be understood that although not shown, other hardware and / or software modules may be used in conjunction with device 12, including but not limited to microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0120] The processing unit 16 executes various functional applications and data processing by running programs stored in the system memory 28, such as implementing the industry-specific power load forecasting method based on federated learning provided in an embodiment of the present invention.

[0121] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A method for forecasting power load by industry based on federated learning, characterized in that: The following steps are involved: Step 1: Construct an industry user dataset based on load influencing factors; Step 2: Establish a load forecasting model based on LSTNet; Step 3: Establish an industry-specific load forecasting model based on federated learning based on the FedAvg algorithm and the FedML framework; Step 3 is to establish an industry-specific load forecasting model based on federated learning based on the FedAvg algorithm and the FedML framework, including the following steps: (1) Let interaction round t = 0, 2, ..., T be the number of local model parameters transmitted by the responding user to the central server located at the active distribution network operator, local training batch e = 1, 2, ..., E be the number of local training iterations of the responding user based on the LSTNet load forecasting model, and model parameter ω be expressed as Equation (18), ω t is the weight coefficient or bias of neurons in each layer of LSTNet after the local training batch reaches E in the tth round of interaction; (2) The central server located at the active distribution network operator sends the initialized model parameters ω0 to all responding users; (3) After the user receives the model parameters, the LSTNet model is trained locally based on the gradient descent algorithm; (4) In the tth round of interaction, respond to user k's update Then it is sent to the central server located at the active distribution network operator; (5) The central server collects all the model parameters updated by the responding users Finally, aggregate the parameters to generate global parameters: Among them, n k is the number of samples of responding user k, and n is the total number of samples of all responding users; (6) The central server sets the global parameter ω t Sent to all responding users, who then update their local parameters based on the global parameters: Then continue training until the Tth interaction is completed.

2. The method for forecasting power load by industry based on federated learning according to claim 1, characterized in that: The factors affecting the load include weather factors, economic factors and time factors.

3. The method for forecasting power load by industry based on federated learning according to claim 1, characterized in that: Before step 1, the original power load data is preprocessed, including the following steps: (1) Use the 3-Sigma criterion to detect outliers in the load data set; (2) Fill in missing values according to formulas (1)-(2): Among them, X is the data value of the original sequence at hour i, x ij is the jth data value in the i-th hour of the interpolated sequence, w ij For X i →x ij The weight of , N is the total number of split points within 1 hour, and T is the total number of hours in the original sequence; (3) Use mean variance normalization to scale features: Among them, x i , μ, σ represent the i-th input data corresponding to a certain feature, the mean of the input data, and the variance of the input data respectively; (4) Set an input sequence sliding window of length P and an output sequence sliding window of length N, and apply these two sliding windows to the time series to obtain an input sequence x of length P. t-1 ,…,x t-P and an output sequence x of length N t ,…,x t+N-1 ; After the input sequence is input into the time series prediction model, the predicted sequence is obtained, as shown in formula (4): Among them, x t represents the input value at time t, represents the predicted value at time t, x t-1 ,...,x t-P represents an input sequence of length P, Represents a prediction sequence of length N; the difference between the output sequence and the prediction sequence is quantified by the evaluation index, which can evaluate the prediction performance of the time series prediction model.

4. The method for forecasting power load by industry based on federated learning according to claim 1, characterized in that: The step 1 of constructing an industry user data set based on load influencing factors includes the following steps: Load, weather and time are selected as the feature types of input data for the load forecasting model; weather data includes temperature, relative humidity, precipitation, wind speed and air pressure; time data includes holiday information, weekday information, year, month, day, hour, minute and day of the week information.

5. The method for power load forecasting by industry based on federated learning according to claim 1 is characterized in that: The step 2 of establishing a load forecasting model based on LSTNet includes the following steps: (1) Convolutional layer setting d c The convolution kernel is of size (m,n), where m is the width of the convolution kernel and n is the number of features. The kth convolution kernel scans the input sequence and outputs h with a length of T. k Vector, as shown in formula (5): h k =RELU(W k *X+b k ) (5) Among them, * is the convolution operation, W k and b k are weight coefficients and bias respectively; the output matrix of the convolutional layer is composed of d c h k Vector composition, size (d c ,T), the corresponding weight coefficients and bias are W C and b C ; (2) The recurrent layer is a GRU model. The unit structure of GRU consists of an update gate and a reset gate. The activation function is RELU. The formulas are shown in (6)-(9): r t =σ(W r [h t-1 ,x t ]+b r ) (6) z t =σ(W z [h t-1 ,x t ]+b z ) (7) c t =RELU(W c [r t *h t-1 ,x t ]+b c ) (8) h t =(1-z t )*h t-1 +z t *c t (9) Among them, r t 、z t and c t are the reset gate coefficient, update gate coefficient and candidate hidden layer coefficient at time t respectively; W r 、W z and W c are the weight coefficients of the reset gate, update gate and candidate hidden layer respectively; b r 、b z and b c are the reset gate, update gate and bias of the candidate hidden layer respectively; h t-1 and h t are the output values of the recurrent layer at time t-1 and time t respectively; (3) The recurrent-jump layer is a GRU model, and p is introduced to capture the periodicity of the time series. The formulas are shown in Equations (10)-(13): r t =σ(W r [h t-p ,x t ]+b r ) (10) z t =σ(W z [h t-p ,x t ]+b z ) (11) c t =RELU(W c [r t *h t-p ,x t ]+b c ) (12) h t =(1-z t )*h t-p +z t *c t (13) Where p is the number of skipped hidden units, h t-p is the output value of the loop-skip layer at time tp; in order to distinguish the loop layer from the loop-skip layer, let is the output value of the recurrent layer at time t, is the output value of the cyclic skip layer at time t; (4) The fully connected layer receives the output of the recurrent layer and the recurrent-skip layer, as shown in formula (14): in, is the output value of the fully connected layer at time t, is the output value of the recurrent-jump layer at time ti, W R and are the weight coefficients of the recurrent layer at time t and the recurrent-skip layer at time i, b D is the bias of the fully connected layer; (5) The autoregressive layer is responsible for providing the linear component, which helps to improve the prediction efficiency in large-scale data; as shown in formula (15): in, is the output of the autoregressive layer, q ar is the size of the input window acting on the input sequence, W ar and b ar are the weight coefficient and bias of the autoregressive layer respectively; (6) The final output prediction result is the superposition of the output of the neural network and the linear component of the autoregressive layer, as shown in formula (16): Among them, y t is the prediction result at time t; the weight coefficient corresponding to the convolution layer is W C , bias is b C , the weight coefficient corresponding to the circulation layer is W R , bias is b R , the weight coefficient corresponding to the cycle-jump layer is W S , bias is b S , the bias of the fully connected layer is b D , the weight coefficient corresponding to the autoregressive layer is W ar , bias is b ar ; The weight coefficients of each layer in LSTNet are W={W C ,W R ,W S ,W ar } and bias b={b C ,b R ,b S ,b D ,b ar } is recorded as: ω∈{W,b}(17) where ω represents the weight coefficient and bias of neurons in each layer, i.e., the model parameters of LSTNet.

6. A device, characterized in that include: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the industry-specific power load forecasting method based on federated learning as described in any one of claims 1-5.

7. A storage medium containing computer-executable instructions, characterized in that: When executed by a computer processor, the computer executable instructions are used to execute the industry-specific power load forecasting method based on federated learning as described in any one of claims 1 to 5.

Images