Deep feature guided two-stage transfer learning for comprehensive energy source-load forecasting

Abstract

This invention discloses a comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning, relating to the field of comprehensive energy multi-element load prediction technology. The method involves acquiring and converting factors such as temperature, season, holidays, wind speed, cloud density, solar intensity, electricity price, and current policies into data formats; acquiring historical data on photovoltaic, wind power, electrical load, heat load, and cooling load in a park, and cleaning the historical data; constructing a time series prediction model based on a deep residual network and a long short-term memory recurrent neural network incorporating an attention mechanism; training the time series prediction model using electrical load, heat load, and cooling load data based on temperature, season, and holidays to obtain a preliminary model; and finally, employing a transfer learning strategy to obtain the final prediction model based on the preliminary model.

Description

Technical Field

[0001] This invention belongs to the field of integrated energy multi-source load forecasting technology, specifically involving an integrated energy source-load forecasting method based on deep feature-guided two-stage transfer learning. Background Technology

[0002] Accurate short-term forecasting of diverse loads is fundamental to the operation and scheduling optimization of regional integrated energy systems, and it is also of great significance for demand-side analysis of the system, such as... Figure 1 As shown, energy service providers for Integrated Energy Systems (IES) in typical areas such as industrial parks, commercial centers, and residential buildings generally need to provide their users with a variety of energy needs, including electricity, heat, and cooling. These needs are significantly influenced by factors such as meteorological conditions, human activities, and building characteristics. Regarding meteorological conditions, due to temperature variations, there are obvious seasonal and regional differences in heating and cooling load demands between the north and south. In terms of human activities, different social behaviors affect the energy consumption characteristics of IES. For example, in residential IES, residents are generally away on weekdays, resulting in mostly rigid loads; while on non-weekdays, residents are more active, leading to flexible and diverse energy-using equipment and random and uncertain energy demands. Different system functions are also an important factor influencing energy consumption characteristics. In industrial areas, electricity load often dominates, with heating and cooling loads playing a supporting role, all conforming to production schedules. In residential areas, electricity and heating loads are often closely related to human activities, exhibiting a certain coupling characteristic between different types of loads. The energy consumption characteristics of IES dictate that changes in one type of energy demand will inevitably lead energy service providers to adjust the other types of energy demand. Load forecasting, as a key aspect of IES... The primary prerequisite for energy demand management and optimized scheduling is that if traditional single load forecasting methods are still used, it is difficult to take into account the differences, randomness, and coupling among different energy demands, and the accuracy of load forecasting cannot be guaranteed. At the same time, given that a large amount of energy conversion coupling information is stored in the IES energy service provider database during long-term operation, it is difficult to extract and summarize the features of these energy conversion characteristics hidden in the data by building detailed mathematical models. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning. The method includes the following steps:

[0004] The source-side and load-side influencing factors of the target park are obtained and converted into data format; wherein, the influencing factors include: temperature, season, holidays, wind speed, cloud density, light intensity, electricity price, gas price and related policies;

[0005] Acquire historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load of the target park within a preset historical time period, and clean the historical data.

[0006] A time series prediction model based on a deep residual network and a long short-term memory recurrent neural network with an attention mechanism is constructed. The time series prediction model is used to predict future time series values ​​based on the changing trends of historical time series data.

[0007] Based on three influencing factors—temperature, season, and holidays—historical data on electrical load, heat load, and cooling load were used to train the time series prediction model to obtain a preliminary model.

[0008] Based on the initial model, six influencing factors are considered: wind speed, cloud density, light intensity, electricity price, gas price, and related policies. Historical data of photovoltaic power generation, wind power generation, electricity load, heat load, and cooling load are used for training to obtain the final photovoltaic power generation prediction model, wind power generation prediction model, electricity load prediction model, heat load prediction model, and cooling load prediction model.

[0009] Preferably, based on the primary model, considering six influencing factors—wind speed, cloud density, solar intensity, electricity price, gas price, and related policies—the model is trained using historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load to obtain the final photovoltaic power generation prediction model, wind power generation prediction model, electrical load prediction model, heat load prediction model, and cooling load prediction model. This includes the following steps:

[0010] Based on the initial model, a transfer learning strategy is adopted, taking into account the influence of light intensity and cloud density, and historical data of photovoltaic power generation is used for transfer training to obtain the final photovoltaic power generation prediction model, so as to predict the photovoltaic output in the future preset time period.

[0011] Based on the initial model, considering the influence of wind speed, historical data of wind turbine power generation is used for transfer training to obtain the final wind turbine power generation prediction model, so as to predict the wind turbine output for a future preset time period.

[0012] Based on the initial model, and considering the influence of relevant policies and electricity prices, the final electricity load prediction model is obtained by using historical electricity load data for transfer training, so as to predict the electricity load for a future preset time period.

[0013] Based on the initial model, considering the influencing factors of electricity and gas prices, the final heat load prediction model is obtained by using historical heat load data for transfer training, so as to predict the heat load for a future preset time period.

[0014] Based on the initial model, and considering the influence of electricity prices, the final cooling load prediction model is obtained by using historical cooling load data for transfer training, in order to predict the cooling load for a future preset time period.

[0015] Preferably, the process of acquiring historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load of the target park within a preset historical time period, and cleaning the historical data, includes the following steps:

[0016] Based on the historical operation information of the target park within a preset historical time period, historical data on photovoltaic power generation, wind turbines, electrical load, heat load, and cooling load of the target park within the preset historical time period are obtained and represented as follows:

[0017]

[0018]

[0019]

[0020]

[0021]

[0022] in, This represents a time series of historical photovoltaic power generation data for the target industrial park. for The power output of photovoltaic power generation at all times. This represents the time series of historical data on wind turbine power generation in the target industrial park. for The power output of the wind turbine at any given time. This represents a time series of historical data on the electrical load of the target industrial park. for Constant electrical load power, This represents a time series of historical data on the cooling load of the target industrial park. for Real-time cooling load data, This represents a time series of historical data on the heat load of the target industrial park. for real-time heat load data, among which ;

[0023] right , , , and For cleaning the sequence data, the window length is first calculated. Local mean of the sequence The formula is as follows:

[0024]

[0025] in As the starting point of the window, P is the position from the starting point of the window. PV(W、E、C、H)_tx+1 for The photovoltaic power generation value, wind turbine power generation value, electrical load power value, cooling load value, and heat load value at any given time.

[0026] Preferably, when there are points in the window that are smaller than the mean, the starting point of the window is replaced with the mean, which has a length of 3 and is centered at the starting point of the window. When there are missing sequences in the window, the mean of the window is used to fill them.

[0027] Preferably, the construction process of the time series prediction model includes the following steps:

[0028] A deep feature extraction network is constructed based on a deep residual network for feature extraction from input data.

[0029] The features extracted by the deep feature extraction network are input into the long short-term memory recurrent neural network based on the attention mechanism to construct a time series prediction model.

[0030] Preferably, the memory unit of the Long Short-Term Memory (LSTM) recurrent neural network is configured by setting an input gate. Output gate And the Gate of Oblivion To select the correction parameters for the error function of memory feedback as a function of gradient descent, in The input data at time is The hidden layer state output value is The memory state is When passing through the forget gate, useless information is discarded to obtain the state of the memory unit at the next moment. Among them, the Gate of Oblivion The calculation formula is as follows:

[0031]

[0032] Among them, h t-1 for The hidden layer state at any given time. and This represents the weight matrix and bias vector in the forget gate. The sigmoid function is used as the activation function.

[0033] The input sequence features are then processed by two activation functions: the sigmoid function and the tanh function, which are used to calculate the data to be input into the memory unit and to create a new candidate state, respectively. The calculation formula is as follows:

[0034]

[0035]

[0036] in, and These are the weight matrix and bias vector in the input gate, respectively. and These are the weight matrix and bias vector in the cell unit state, respectively;

[0037] Unit state value at time t Hidden layer state at time t and output value The calculation formulas are as follows:

[0038]

[0039]

[0040]

[0041] Among them, c t-1 The element state value at time t-1 For activation function, and These are the weight matrix and bias vector in the output gate, respectively.

[0042] Preferably, the construction process of a long short-term memory recurrent neural network based on an attention mechanism includes the following steps:

[0043] Introducing an attention mechanism into a Long Short-Term Memory (LSTM) recurrent neural network to control the hidden layer states output by the LSM hidden layer. Perform attention weight allocation, using To represent the attention probability distribution values, the attention weight matrix and eigenvectors The calculation formula is as follows:

[0044]

[0045]

[0046]

[0047] in, , For the length of the input data, For attention weights, for For the current input The assigned attention weights for Unnormalized weight matrix, , and These are the randomly initialized Attention weight matrix, bias, and time series matrix, respectively.

[0048] On the other hand, a comprehensive energy source-load prediction device with deep feature-guided two-stage transfer learning is also provided, comprising:

[0049] The data acquisition module is used to acquire the source-side and load-side influencing factors of the target park and convert them into data format; wherein, the influencing factors include: temperature, season, holidays, wind speed, cloud density, light intensity, electricity price, gas price and related policies;

[0050] The data processing module is used to acquire historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load and cooling load of the target park within a preset historical time period, and to clean the historical data.

[0051] The first building module is used to construct a time series prediction model based on a deep residual network and a long short-term memory recurrent neural network with an attention mechanism. The time series prediction model is used to predict future time series values ​​based on the changing trends of historical time series data.

[0052] The second building module is used to train the time series prediction model based on three influencing factors: temperature, season, and holidays, using historical data of electrical load, heat load, and cooling load to obtain a primary model.

[0053] The third module is used to build upon the primary model by considering six influencing factors: wind speed, cloud density, light intensity, electricity price, gas price, and related policies. It uses historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load for training to obtain the final photovoltaic power generation prediction model, wind power generation prediction model, electrical load prediction model, heat load prediction model, and cooling load prediction model.

[0054] On the other hand, a device is also provided, comprising:

[0055] One or more processors;

[0056] Memory, used to store one or more computer programs;

[0057] When the one or more computer programs are executed by the one or more processors, the one or more processors implement a comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning as described above.

[0058] On the other hand, a storage medium is also provided, on which a computer program is stored, which, when executed by a processor, is used to perform a comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning as described above.

[0059] The beneficial effects of this invention are:

[0060] This invention employs a two-stage training process. It obtains a preliminary model by considering common factors across multiple influencing factors, and then uses transfer learning based on specific factors to obtain a multivariate load prediction model. This improves the model's training efficiency and generalization ability. The transition from comprehensive learning in the first stage to precise learning in the second stage significantly enhances the model's prediction accuracy. Based on multivariate load prediction data, this invention enables energy service providers to accurately predict multivariate load demands by aggregating and analyzing various energy needs of different users. It also coordinates the conversion, storage, distribution, and consumption processes within the energy service provider ecosystem (IES) to meet the diverse energy needs of different users. Furthermore, by using an Attention-based LSTM model, attention is focused on key areas at specific moments, reducing or even ignoring attention to other areas. This allows for the acquisition of more relevant details and the suppression of useless information, resulting in better multivariate prediction performance and significantly improved prediction accuracy. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a simplified model diagram of the integrated energy interaction structure of the present invention;

[0063] Figure 2 This is a block diagram of a two-stage integrated energy source-load prediction model according to an embodiment of the present invention;

[0064] Figure 3 This is the sequence processing model structure of LSTM according to an embodiment of the present invention;

[0065] Figure 4 This is a schematic diagram of the LSTM-based attention mechanism in an embodiment of the present invention;

[0066] Figure 5 This is a schematic diagram of the hotel load prediction results according to an embodiment of the present invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] like Figure 2 As shown, a comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning includes the following steps:

[0069] The influencing factors on the source and load sides of the target industrial park are obtained and converted into data format. These influencing factors include: temperature, season, holidays, wind speed, cloud density, solar intensity, electricity price, gas price, and relevant policies. Temperature, season, and holidays, as common influencing factors, all have a certain impact on the source and load of the industrial park. Wind speed is highly correlated with wind turbine power generation. Cloud density and solar intensity are highly correlated with photovoltaic power generation. Electricity price and current policies are highly correlated with electricity load. Electricity price directly affects the total electricity load, while current policies, such as the "dual-carbon" strategy and the orderly production plan issued by the State Grid, affect the total regional electricity load. Electricity price and gas price are highly correlated with heat load, as heat load is partly generated by electricity and gas, and electricity and gas prices affect the demand and output of heat load. Electricity price is also highly correlated with cooling load, as cooling load mainly comes from electricity production, and electricity price also affects the generation of cooling load. Excessively high electricity prices may cause some users to shut down cooling load generating devices such as air conditioners.

[0070] Acquire historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load of the target park within a preset historical time period, and clean the historical data.

[0071] A time series prediction model based on a deep residual network and a long short-term memory recurrent neural network with an attention mechanism is constructed. The time series prediction model is used to predict future time series values ​​based on the changing trends of historical time series data.

[0072] By using three factors—temperature, season, and holidays—and electrical load, heat load, and cooling load data, a time series forecasting model is trained to obtain a preliminary model, which is used to predict future series values ​​based on the changing trends of historical data.

[0073] Based on the initial model, six influencing factors are considered: wind speed, cloud density, light intensity, electricity price, gas price, and related policies. Historical data of photovoltaic power generation, wind power generation, electricity load, heat load, and cooling load are used for training to obtain the final photovoltaic power generation prediction model, wind power generation prediction model, electricity load prediction model, heat load prediction model, and cooling load prediction model.

[0074] Since photovoltaic power generation will be used for the park's electrical load and heating and cooling load, there is an inherent coupling relationship. Therefore, based on the primary model, a transfer learning strategy is adopted to use photovoltaic data for transfer training to obtain the final photovoltaic power generation prediction model. This model is used to predict future photovoltaic output based on historical power data, providing data support for the subsequent participation of photovoltaic in the energy dispatch of integrated energy.

[0075] Since wind power generation will be used for the park's electrical load and heating and cooling load, there is an inherent coupling relationship. Therefore, based on the primary model, wind power data is used for transfer training to obtain the final wind power generation prediction model. This model is used to predict the future output of wind turbines based on historical power generation data, providing data support for the subsequent participation of wind turbines in the energy dispatch of integrated energy.

[0076] Based on the initial model, the final electricity load prediction model is obtained by using electricity load data for transfer training. This model is used to predict future electricity load data based on historical electricity consumption data, providing electricity load data support for subsequent integrated energy dispatch and optimized operation.

[0077] Based on the initial model, the heat load data is used for transfer training to obtain the final heat load prediction model, which is used to predict the future heat load from historical heat load data, providing heat load data support for subsequent integrated energy dispatch and optimized operation.

[0078] Based on the initial model, the final cooling load prediction model is obtained by using cooling load data for transfer training. This model is used to predict future cooling loads based on historical cooling load data, providing cooling load data support for subsequent integrated energy dispatch and optimized operation.

[0079] It should be further explained that, in the specific implementation process, the influencing factors include temperature. ,season Holidays Wind speed Cloud density Light intensity Electricity price And policy information.

[0080] It should be further explained that, in the specific implementation process, the process of acquiring historical data on the park's photovoltaic, wind power, electrical load, heat load, and cooling load, and cleaning the historical data, includes the following steps:

[0081] Based on the historical operation information of the target park within a preset historical time period, historical data on photovoltaic power generation, wind turbines, electrical load, heat load, and cooling load of the target park within the preset historical time period are obtained and represented as follows:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] in, This represents a time series of historical photovoltaic power generation data for the target industrial park. for The power output of photovoltaic power generation at all times. This represents the time series of historical data on wind turbine power generation in the target industrial park. for The power output of the wind turbine at any given time. This represents a time series of historical data on the electrical load of the target industrial park. for Constant electrical load power, This represents a time series of historical data on the cooling load of the target industrial park. for Real-time cooling load data, This represents a time series of historical data on the heat load of the target industrial park. for real-time heat load data, among which ;

[0088] right , , , and For cleaning the sequence data, the window length is first calculated. Local mean of the sequence P pv（W、E、C、H） The formula is as follows:

[0089]

[0090] in As the starting point of the window, P is the position from the start of the window. PV(W、E、C、H)_tx+1 for The photovoltaic power generation value, wind turbine power generation value, electrical load power value, cooling load value, and heat load value at any given time.

[0091] It should be further explained that, in the specific implementation process, when there are points in the window that are smaller than the mean, the starting point of the window is replaced with the mean, which has a length of 3 and is centered at the starting point of the window. When there are missing sequences in the window, the mean of the window is used to fill them.

[0092] It should be further explained that, in the specific implementation process, the construction process of the time series prediction model includes the following steps:

[0093] A deep feature extraction network is constructed based on a deep residual network for feature extraction from input data.

[0094] The features extracted by the deep residual network are input into a Long Short-Term Memory (LSTM) recurrent neural network based on the attention mechanism to construct a time series prediction model. The LSTM sequence processing model structure is as follows: Figure 3 As shown, the attention-based long short-term memory recurrent neural network is as follows: Figure 4 As shown.

[0095] It should be further explained that, in specific implementation, the LSTM recurrent neural network is an improved version of the recurrent neural network (RNN). While maintaining the basic structure of the RNN, the LSTM recurrent neural network redesigns the memory units and sets up input gates. Output gate And the Gate of Oblivion To select the correction parameters for the error function of memory feedback as a function of gradient descent, in The input data at time is The hidden layer state output value is The memory state is When passing through the forget gate, useless information is discarded to obtain the state of the memory unit at the next moment. Among them, the Gate of Oblivion The calculation formula is as follows:

[0096]

[0097] Among them, h t-1 for The hidden layer state at any given time. and This represents the weight matrix and bias vector in the forget gate. The activation function is the sigmoid function, and the output is... The value of is between 0 and 1, and will be successively ANDed with . When elements are multiplied, When a bit has a value of 0, the information of the corresponding bit is completely discarded. When the value is (0, 1), part of the information of the corresponding bit is retained and part is discarded. Only when the value is 1 will the information of the corresponding bit be completely retained.

[0098] The input sequence features are then processed by two activation functions: the sigmoid function and the tanh function, which are used to calculate the data to be input into the memory unit and to create a new candidate state, respectively. The calculation formula is as follows:

[0099]

[0100]

[0101] in, and These are the weight matrix and bias vector in the input gate, respectively. and These are the weight matrix and bias vector in the cell unit state, respectively;

[0102] Unit state value at time t Hidden layer state at time t and output value The calculation formulas are as follows:

[0103]

[0104]

[0105]

[0106] Among them, c t-1 The element state value at time t-1 For activation function, and These are the weight matrix and bias vector in the output gate, respectively.

[0107] It should be further explained that, in the specific implementation process, the Attention mechanism can allocate more attention to the key parts of the input sequence that affect the output result, thereby better learning the information in the sequence. Introducing the Attention mechanism into LSTM affects the output values ​​of the LSTM hidden layers. Perform attention weight allocation, using To represent the attention probability distribution values, the attention weight matrix and eigenvectors The calculation formula is as follows:

[0108]

[0109]

[0110]

[0111] in, , For the length of the input data, For attention weights, for For the current input The assigned attention weights The weight matrix is ​​unnormalized. , and These are the randomly initialized Attention weight matrix, bias, and time series matrix, respectively.

[0112] It should be further explained that, in the specific implementation process, the final prediction models for photovoltaic, wind power, electrical load, heat load and cooling load are obtained by adopting a transfer learning strategy based on the primary time series prediction model.

[0113] For photovoltaic power generation at the source, the effects of light intensity and cloud density are further considered, and the photovoltaic data is used for transfer training to obtain the final photovoltaic power generation prediction model.

[0114] To further consider the impact of wind speed on the source side of wind power, the final wind power generation prediction model is obtained by using wind power data for transfer training.

[0115] To further consider the impact of policy guidance and electricity prices on the load side, the final load prediction model is obtained by using load data for transfer training.

[0116] To further consider the impact of electricity and gas prices on the heat load side, the heat load data was used for transfer training to obtain the final heat load prediction model.

[0117] To further consider the impact of electricity prices on the load-side cooling load, the final cooling load prediction model is obtained by using cooling load data for transfer learning.

[0118] Example: To verify the effectiveness of the proposed method, taking an electricity load forecasting model as an example, 35,040 data points were obtained from a hotel from 00:00 on January 1, 2021 to 24:00 on December 31, 2021, at a sampling frequency of 15 minutes. In the neural network, the training set used data from 00:00 on January 1, 2021 to 24:00 on December 30, 2021, and the test set used data from 00:15 on December 31, 2021 to 24:00 on December 31, 2021, totaling 24 hours and 96 points. The LSTM layer trains the reconstructed data in a "many-to-one" pattern, using 64 points as input and the next point as output, providing the prediction results as follows. Figure 5 As shown:

[0119] The dashed line represents the load forecast curve, and the solid line represents the actual load curve. The error evaluation index is defined as follows:

[0120]

[0121] in n This represents the sample size of the test set. The actual load sequence for the test set; Predict the load sequence for the test set. MAPE is a good measure of the average magnitude of the prediction error. Figure 5 It can be seen that the predicted data and the actual load are basically consistent, with an error index (MAPE) of 0.032. In summary, the effectiveness of the proposed integrated energy source-load prediction method based on deep feature-guided two-stage transfer learning is evident.

[0122] It should be further explained that, based on the same inventive concept, this invention also provides a comprehensive energy source-load prediction device guided by deep features for two-stage transfer learning, characterized in that the method includes the following steps:

[0123] The data acquisition module is used to acquire the source-side and load-side influencing factors of the target park and convert them into data format; wherein, the influencing factors include: temperature, season, holidays, wind speed, cloud density, light intensity, electricity price, gas price and related policies;

[0124] The data processing module is used to acquire historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load and cooling load of the target park within a preset historical time period, and to clean the historical data.

[0125] The first building module is used to construct a time series prediction model based on a deep residual network and a long short-term memory recurrent neural network with an attention mechanism. The time series prediction model is used to predict future time series values ​​based on the changing trends of historical time series data.

[0126] The second building module is used to train the time series prediction model based on three influencing factors: temperature, season, and holidays, using historical data of electrical load, heat load, and cooling load to obtain a primary model.

[0127] The third module is used to build upon the primary model by considering six influencing factors: wind speed, cloud density, light intensity, electricity price, gas price, and related policies. It uses historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load for training to obtain the final photovoltaic power generation prediction model, wind power generation prediction model, electrical load prediction model, heat load prediction model, and cooling load prediction model.

[0128] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0129] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0130] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0131] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A comprehensive energy source-load prediction method guided by deep features and employing two-stage transfer learning, characterized in that... The method includes the following steps: The source-side and load-side influencing factors of the target park are obtained and converted into data format; wherein, the influencing factors include: temperature, season, holidays, wind speed, cloud density, light intensity, electricity price, gas price and related policies; Acquire historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load of the target park within a preset historical time period, and clean the historical data. A time series prediction model based on a deep residual network and a long short-term memory recurrent neural network with an attention mechanism is constructed. The time series prediction model is used to predict future time series values ​​based on the changing trends of historical time series data. Based on three influencing factors—temperature, season, and holidays—historical data on electrical load, heat load, and cooling load were used to train the time series prediction model to obtain a preliminary model. Based on the initial model, six influencing factors are considered: wind speed, cloud density, light intensity, electricity price, gas price, and related policies. Historical data of photovoltaic power generation, wind power generation, electricity load, heat load, and cooling load are used for training to obtain the final photovoltaic power generation prediction model, wind power generation prediction model, electricity load prediction model, heat load prediction model, and cooling load prediction model.

2. The comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning according to claim 1, characterized in that, Based on the initial model, considering six influencing factors—wind speed, cloud density, solar irradiance, electricity price, gas price, and related policies—historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load are used for training to obtain the final photovoltaic power generation prediction model, wind power generation prediction model, electrical load prediction model, heat load prediction model, and cooling load prediction model. This process includes the following steps: Based on the initial model, a transfer learning strategy is adopted, taking into account the influence of light intensity and cloud density, and historical data of photovoltaic power generation is used for transfer training to obtain the final photovoltaic power generation prediction model, so as to predict the photovoltaic output in the future preset time period. Based on the initial model, considering the influence of wind speed, the final wind turbine power generation prediction model is obtained by using historical data of wind turbine power generation for transfer training, so as to predict the wind turbine output for a future preset time period. Based on the initial model, and considering the influence of relevant policies and electricity prices, the final electricity load prediction model is obtained by using historical electricity load data for transfer training, so as to predict the electricity load for a future preset time period. Based on the initial model, considering the influencing factors of electricity and gas prices, the final heat load prediction model is obtained by using historical heat load data for transfer training, so as to predict the heat load for a future preset time period. Based on the initial model, and considering the influence of electricity prices, the model is trained using historical data on cooling load to obtain the final cooling load prediction model, which is used to predict the cooling load for a future preset time period.

3. The comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning according to claim 1, characterized in that, The process of acquiring historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load of the target park within a preset historical time period, and cleaning the historical data, includes the following steps: Based on the historical operation information of the target park within a preset historical time period, historical data on photovoltaic power generation, wind turbines, electrical load, heat load, and cooling load of the target park within the preset historical time period are obtained and represented as follows: in, This represents the time series of historical photovoltaic power generation data for the target industrial park. for The power output of photovoltaic power generation at any given time This represents the time series of historical data on wind turbine power generation in the target industrial park. for The power output of the wind turbine at any given time. This represents a time series of historical data on the electrical load of the target industrial park. for Constant electrical load power, This represents a time series of historical data on the cooling load of the target industrial park. for Real-time cooling load data, This represents a time series of historical data on the heat load of the target industrial park. for real-time heat load data, among which ; right , , , and For cleaning the sequence data, the window length is first calculated. Local mean of the sequence The formula is as follows: in As the starting point of the window, This is the position relative to the starting point of the window. for The photovoltaic power generation value, wind turbine power generation value, electrical load power value, cooling load value, and heat load value at any given time.

4. The comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning according to claim 3, characterized in that, When there are points in the window that are smaller than the mean, the starting point of the window is replaced with the mean, which is 3 points long and has the starting point as its midpoint. When there are missing points in the sequence, the mean of the window is used to fill them.

5. The comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning according to claim 1, characterized in that, The construction process of the time series prediction model includes the following steps: A deep feature extraction network is constructed based on a deep residual network for feature extraction from input data. The features extracted by the deep feature extraction network are input into the long short-term memory recurrent neural network based on the attention mechanism to construct a time series prediction model.

6. The comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning according to claim 5, characterized in that, The memory units of a Long Short-Term Memory (LSTM) recurrent neural network are configured by setting input gates. Output gate And the Gate of Oblivion To select the correction parameters for the error function of memory feedback as a function of gradient descent, in The input data at time is The hidden layer state output value is The memory state is When passing through the forget gate, useless information is discarded to obtain the state of the memory unit at the next moment. Among them, the Gate of Oblivion The calculation formula is as follows: in, for The hidden layer state at any given time. and This represents the weight matrix and bias vector in the forget gate. The sigmoid function is used as the activation function. The input sequence features are then processed by two activation functions: the sigmoid function and the tanh function, which are used to calculate the data to be input into the memory unit and to create a new candidate state, respectively. The calculation formula is as follows: in, and These are the weight matrix and bias vector in the input gate, respectively. and These are the weight matrix and bias vector in the cell unit state, respectively; Unit state value at time t Hidden layer state at time t and output value The calculation formulas are as follows: in, The element state value at time t-1 For activation function, and These are the weight matrix and bias vector in the output gate, respectively.

7. The comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning according to claim 6, characterized in that, The construction process of a long short-term memory recurrent neural network based on the attention mechanism includes the following steps: Introducing an attention mechanism into a Long Short-Term Memory (LSTM) recurrent neural network to control the hidden layer states output by the LSM hidden layer. Attention weights are assigned, and an attention weight matrix is ​​generated. and eigenvectors The calculation formula is as follows: in, , For the length of the input data, For attention weights, For the current input The assigned attention weights for Unnormalized weight matrix, , and These are the randomly initialized Attention weight matrix, bias, and time series matrix, respectively.

8. A comprehensive energy source-load prediction device based on deep feature-guided two-stage transfer learning, characterized in that, include: The data acquisition module is used to acquire the source-side and load-side influencing factors of the target park and convert them into data format; wherein, the influencing factors include: temperature, season, holidays, wind speed, cloud density, light intensity, electricity price, gas price and related policies; The data processing module is used to acquire historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load and cooling load of the target park within a preset historical time period, and to clean the historical data. The first building module is used to construct a time series prediction model based on a deep residual network and a long short-term memory recurrent neural network with an attention mechanism. The time series prediction model is used to predict future time series values ​​based on the changing trends of historical time series data. The second building module is used to train the time series prediction model based on three influencing factors: temperature, season, and holidays, using historical data of electrical load, heat load, and cooling load to obtain a primary model. The third module is used to build upon the primary model by considering six influencing factors: wind speed, cloud density, light intensity, electricity price, gas price, and related policies. It uses historical data on photovoltaic power generation, wind turbine power generation, electrical load, heat load, and cooling load for training to obtain the final photovoltaic power generation prediction model, wind power generation prediction model, electrical load prediction model, heat load prediction model, and cooling load prediction model.

9. A device, characterized in that, include: One or more processors; Memory, used to store one or more computer programs; When the one or more computer programs are executed by the one or more processors, the one or more processors implement the comprehensive energy source-load prediction method of deep feature-guided two-stage transfer learning as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, is used to perform the comprehensive energy source-load prediction method based on deep feature-guided two-stage transfer learning as described in any one of claims 1-7.

Images