Power load prediction method and system based on deep learning

Abstract

The invention discloses a power load prediction method and system based on deep learning, and the method comprises the steps: inputting a second original data sequence into a first RNN network layer of a pre-constructed power load prediction model, and inputting a first target original data sequence into a first MLP network layer of the power load prediction model; performing external weight amplification on the output result of the first RNN network to obtain a first output result, transmitting the output result of the first MLP network layer and the first output result to a second MLP network layer of the power load prediction model, and outputting a third output result by the second MLP network layer; and inputting the third output result into at least one second RNN network layer of the power load prediction model, and outputting by a certain second RNN network layer to obtain a load prediction result. The load can be accurately predicted in different prediction time periods, and demand planning can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of load forecasting, and in particular relates to a method and system for power load forecasting based on deep learning. Background Art

[0002] Power load forecasting is one of the important tasks of the power sector, providing key auxiliary functions for the formulation and implementation of national energy layout and economic development strategies. In particular, medium- and long-term power load forecasting is of great significance for power grid planning. However, the current medium- and long-term load forecasting has problems such as difficulty in ensuring model accuracy, slow technology upgrades, and lack of cutting-edge technology exploration.

[0003] Solving the above problems is crucial to improving the scientificity and reliability of power grid planning and dispatching. However, the economic development level, climate conditions, population structure and electricity consumption habits of different regions vary, making it difficult to directly migrate the effective load forecasting model in one region to other regions. Traditional physical models built based on domain knowledge require detailed investigation of the power grid operation mode and analysis of the physical data relationships of a large number of related inputs. The model building cycle is long and the cost is high, which makes it difficult to meet the actual application needs. Summary of the invention

[0004] The present invention provides a method and system for power load forecasting based on deep learning, which are used to solve the technical problems of needing to analyze the physical data relationship of a large number of related inputs, and having a long model building cycle and high cost.

[0005] In a first aspect, the present invention provides a method for power load forecasting based on deep learning, comprising:

[0006] Acquire a first original data sequence uploaded by the central control platform within a first historical time period, and perform data filling on the first original data sequence using a preset missing value filling rule to obtain a first target original data sequence;

[0007] Acquire a second original data sequence reinforced with input in a second historical time period, wherein the second historical time period does not overlap with the first historical time period, and the second historical time period is a historical time period adjacent to the current moment;

[0008] The second original data sequence is input into the first RNN network layer of the pre-built power load forecasting model, and the first target original data sequence is input into the first MLP network layer of the power load forecasting model, wherein the power load forecasting model is obtained through iterative training of the neural network, and the expression of the loss function of the iterative training of the neural network is:

[0009]

[0010] In the formula, is the penalty term, a is the weight coefficient, k is the number of predicted samples with negative values, w and b are self-optimizing parameters, n is the number of samples, and x k is the input vector when the predicted value is negative, and x is the input vector of the sample;

[0011] Performing external weight amplification on the output result of the first RNN network to obtain a first output result, and transmitting the output result of the first MLP network layer and the first output result to the second MLP network layer of the power load forecasting model, and the second MLP network layer outputs a third output result;

[0012] The third output result is input into at least one second RNN network layer of the power load forecasting model, and a second RNN network layer outputs a first target load forecasting result of the first level at the current moment, a second target load forecasting result of the second level at the current moment, and a third target load forecasting result of the third level at the current moment;

[0013] Based on a preset data fusion device, the first target load prediction result is fused with the third target load prediction result to obtain a first fusion result, and the second target load prediction result is fused with the third target load prediction result to obtain a second fusion result;

[0014] A preset neural network model is used to determine the nonlinear relationship between the first fusion result and the second fusion result, and the third target load forecast result is corrected according to the nonlinear relationship to obtain the final load forecast result of the third level at the current moment. The expression of the final load forecast result is: In the formula, is the first fusion result, is the second fusion result, and W is the parameter of the neural network model.

[0015] In a second aspect, the present invention provides a power load forecasting system based on deep learning, comprising:

[0016] a filling module, configured to obtain a first original data sequence uploaded by the central control platform within a first historical time period, and perform data filling on the first original data sequence using a preset missing value filling rule to obtain a first target original data sequence;

[0017] an acquisition module, configured to acquire a second original data sequence reinforced by input in a second historical time period, wherein the second historical time period does not overlap with the first historical time period, and the second historical time period is a historical time period adjacent to the current moment;

[0018] The input module is configured to input the second original data sequence into the first RNN network layer of the pre-constructed power load forecasting model, and input the first target original data sequence into the first MLP network layer of the power load forecasting model, wherein the power load forecasting model is obtained through iterative training of the neural network, and the expression of the loss function of the iterative training of the neural network is:

[0019]

[0020] In the formula, is the penalty term, a is the weight coefficient, k is the number of predicted samples with negative values, w and b are self-optimizing parameters, n is the number of samples, and x k is the input vector when the predicted value is negative, and x is the input vector of the sample;

[0021] A processing module is configured to perform external weight amplification on the output result of the first RNN network to obtain a first output result, and transmit the output result of the first MLP network layer and the first output result to the second MLP network layer of the power load forecasting model, and the second MLP network layer outputs a third output result;

[0022] An output module is configured to input the third output result into at least one second RNN network layer of the power load forecasting model, and a second RNN network layer outputs a first target load forecast result of the first level at the current moment, a second target load forecast result of the second level at the current moment, and a third target load forecast result of the third level at the current moment;

[0023] a fusion module, configured to fuse the first target load prediction result with the third target load prediction result based on a preset data fusion device to obtain a first fusion result, and to fuse the second target load prediction result with the third target load prediction result to obtain a second fusion result;

[0024] The correction module is configured to use a preset neural network model to determine the nonlinear relationship between the first fusion result and the second fusion result, and to correct the third target load forecast result according to the nonlinear relationship to obtain a final load forecast result of the third level at the current moment, wherein the expression of the final load forecast result is: In the formula, is the first fusion result, is the second fusion result, and W is the parameter of the neural network model.

[0025] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the deep learning-based power load forecasting method of any embodiment of the present invention.

[0026] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the program instructions are executed by a processor, the processor executes the steps of the deep learning-based power load forecasting method of any embodiment of the present invention.

[0027] The power load forecasting method and system based on deep learning in this application builds a multi-data input fusion short-term forecasting model combining traditional machine learning and deep learning neural networks based on the historical time series collected by the central control platform and the valid data after exogenous variable expansion as input. It can accurately forecast the load in different forecast time periods and realize demand planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 A flowchart of a method for power load forecasting based on deep learning provided by one embodiment of the present invention;

[0030] Figure 2 A schematic diagram of the structure of a power load prediction model of a specific embodiment is provided for an embodiment of the present invention;

[0031] Figure 3 A histogram of absolute value error distribution of an experimental model and a control group model of a specific embodiment is provided for one embodiment of the present invention;

[0032] Figure 4 A structural block diagram of a power load forecasting system based on deep learning provided by an embodiment of the present invention;

[0033] Figure 5 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] See also Figure 1 , which shows a flow chart of a power load forecasting method based on deep learning in the present application.

[0036] like Figure 1 As shown, the power load forecasting method based on deep learning specifically includes the following steps:

[0037] Step S101: obtain a first original data sequence uploaded by the central control platform within a first historical time period, and perform data filling on the first original data sequence using a preset missing value filling rule to obtain a first target original data sequence.

[0038] In this step, the first original data sequence uploaded by the central control platform in the first historical time period is obtained, and the missing attributes are interpolated multiple times. In each iteration, the random forest is used to predict the value close to the missing value to replace it, and finally the complete first original data sequence is obtained.

[0039] Specifically, the first step is data collection and preparation. From January 1, 2022 to December 31, 2023, a total of 730 time points were collected at intervals of 1 day. At the same time, local meteorological data related to the load were collected: temperature, humidity, wind speed and solar radiation. A major advantage of the data-driven algorithm based on deep learning is that it analyzes the functional mapping relationship of a single input to the prediction target in a disordered and detailed manner. The algorithm will perform data learning on its own to assign the corresponding weight to each input. For some missing data, multiple interpolation is performed using chain equations based on random forests. By using chain equations for optimization based on the multiple interpolation method, some missing values ​​are randomly selected for interpolation, and the interpolation data set is gradually iterated and optimized. The specific steps include: first, obtaining the initial random interpolation data set, and then performing multiple iterative interpolation of the missing attributes. In each iteration, the random forest is used to predict the value close to the missing value to replace it, and finally a complete data set is obtained. This method continuously optimizes the interpolation results, and finally realizes multiple interpolation of missing data, enhancing the integrity and accuracy of the data.

[0040] Step S102, obtaining a second original data sequence reinforced by input in a second historical time period, wherein the second historical time period does not overlap with the first historical time period, and the second historical time period is a historical time period adjacent to the current moment.

[0041] In this step, according to the load trend, the measured values ​​of the previous k minutes of load history (5 is selected based on experiments in this project, and it can also be up to 15. If it is greater than 15, the effect is limited, but the computing complexity increases) are used as enhanced input, which can effectively improve the accuracy of the algorithm by 10-15%.

[0042] Step S103: input the second original data sequence into the first RNN network layer of the pre-built power load forecasting model, and input the first target original data sequence into the first MLP network layer of the power load forecasting model.

[0043] In this step, the neural network is calculated by connecting a large number of artificial neurons. In most cases, artificial neural networks can change their internal structure based on external information and are adaptive systems. A typical neural network has the following three key parts:

[0044] Architecture refers to the variables in the network and their topological relationships. For example, the number of hidden layers and the number of neurons in each layer in a neural network.

[0045] Activation function: Most neural network models have a short-time-scale dynamics rule that defines how neurons change their activation values ​​based on the activity of other neurons.

[0046] Learning rate refers to how the weights in the network are adjusted over time. This is generally regarded as a long-term dynamic rule.

[0047] The basic building block of a neural network is a neuron. The input x received by each neuron is multiplied by a weight w, then summed and added with a bias value b. An activation function f is applied to the sum to obtain the final output y. The expression of the output y is:

[0048] y=f(Σwx+b),

[0049] The activation function expression of the hidden layer in the neural network is:

[0050]

[0051] Where f(z) is the tanh function;

[0052] The activation function expression of the output layer in the neural network is:

[0053]

[0054] Where f(z)′ is the sigmoid function.

[0055] The weight w and the bias b are obtained by the model through the back propagation algorithm based on the historical data. The mainstream supervised learning method is adopted in the present invention. By preparing the data with results / labels (the load value at the future moment), the model is allowed to establish a model on the training data set through the back propagation algorithm (optimizing and fixing the internal parameters, i.e., weights and biases). Then, when new data appears in the future practical application (no results / labels at this time), it can accept input and infer new instances based on the trained model. The training optimizer used in this invention is Adam. The hyperparameters of the neural network model used in the present invention, such as the number of hidden layers, the number of neurons in each layer, the learning rate, etc., are obtained and optimized through evolutionary algorithms.

[0056] The evolutionary algorithm is inspired by the biological evolution mechanism and simulates the evolutionary process of natural selection for model optimization. The evolutionary algorithm searches for a better solution in the search space through multiple rounds of iterative processes. The process of the evolutionary algorithm is as follows: In the initial round, the evolutionary algorithm randomly generates a batch of individuals. The parameters to be optimized owned by these individuals are uniformly selected from the state space. Then the evolutionary algorithm evaluates the fitness of each individual. In the present invention, the fitness is the prediction error, and a higher fitness corresponds to a lower prediction error. When the fitness calculation of all individuals is completed, the winning individuals will generate the next generation by crossover and mutation. The next generation generated uses the same defined fitness to evaluate performance. The above process is repeated through multiple rounds until the overall average fitness converges. For more details on the implementation of the evolutionary algorithm, please refer to the literature (Y.Chu, HTCPedro, CFMCoimbra, Hybrid intra-hour DNI forecasts with sky image processing enhanced by stochastic learning, Sol.Energy.98 (2013) 592-603.). The evolutionary algorithm requires high computing power. The evaluation of all individuals in each generation and the calculation of the next generation require about 16 computing cores for 1 hour. The overall optimization requires about 50 generations of individuals, which takes about 2 days of computing time in total. In the experiment of the present invention, the final optimization result is a double hidden layer, each layer has 10 neurons, and the learning rate is 0.02.

[0057] Specifically, an original data sequence within a preset historical time period is obtained, and the original data sequence is filled with data using a preset missing value filling rule to obtain a target original data sequence, wherein the target original data sequence includes a target original data training subsequence and a target original data test subsequence; the target original data training subsequence is input into a neural network for iterative training to obtain an initial power load forecasting model, and the initial power load forecasting model is tested based on the target original data test subsequence to obtain a final power load forecasting model, wherein the power load forecasting model includes an inverse K-type network structure, such as Figure 2 As shown, the inverse K-type network structure includes a first RNN network layer, a first MLP network layer, a second MLP network layer and at least one second RNN network layer.

[0058] For the newly added enhanced input part in the second part, it is found that it has a greater positive effect on the prediction results. A special inverse k-type network structure is designed to amplify its influence in the network through 100-1000 times of external weights. The structure of the new enhanced input part is as follows. In addition, because the enhanced input is a historical time series, it is not connected to the network for input processing of the central control platform. It is an RNN dedicated to analyzing time series (LSTM or GRU can also be used). This specific structure can well consider the fusion of historical trends and current key information, so as to effectively predict the trend of future load. At the output end, if the prediction target is a future time series, a RNN prediction head is added. In addition, another major advantage of this model is that it can be used to predict other related indicators that need to be controlled in addition to load, such as temperature, humidity, etc. These prediction objects share the same prediction input, which can effectively reduce the complexity of the model and the amount of inference calculation. The specific method is the same as the load prediction. Add another prediction head to the second hidden layer. If it is a time series prediction, add an RNN head, if it is a single value, use a regular neuron.

[0059] It should be noted that due to the randomness of the data-driven algorithm, negative values ​​that do not conform to the actual physical laws may occur. Therefore, two methods are combined. One is to add a penalty term to the loss function of the model training. The expression of the loss function of the iterative training of the neural network is:

[0060]

[0061] In the formula, is the penalty term, a is the weight coefficient, k is the number of predicted samples with negative values, w and b are self-optimizing parameters, n is the number of samples, and x k is the input vector when the predicted value is negative, and x is the input vector of the sample;

[0062] In addition, during the inference phase, if extreme cases with negative values ​​still occur, an external model is used for clipping max(0, f(x k , w, b).

[0063] Step S104, externally weighting the output result of the first RNN network to obtain a first output result, and transmitting the output result of the first MLP network layer and the first output result to the second MLP network layer of the power load forecasting model, and the second MLP network layer outputs a third output result.

[0064] Step S105, input the third output result into at least one second RNN network layer of the power load forecasting model, and a second RNN network layer outputs the first target load forecasting result of the first level at the current moment, the second target load forecasting result of the second level at the current moment, and the third target load forecasting result of the third level at the current moment.

[0065] In this step, the first level, the second level, and the third level are minute level, hour level, and day level, respectively.

[0066] Step S106: Based on a preset data fusion device, the first target load prediction result and the third target load prediction result are fused to obtain a first fusion result, and the second target load prediction result and the third target load prediction result are fused to obtain a second fusion result.

[0067] In this step, the expression of the first fusion result is:

[0068]

[0069] In the formula, is the first fusion result, is the first target load forecast result, is the power load characteristic of the third level at the current moment, is the latitude of the feature, is the third target load prediction result;

[0070] The expression of the second fusion result is:

[0071]

[0072] In the formula, is the second fusion result, This is the second target load prediction result.

[0073] Step S107, using a preset neural network model to determine the nonlinear relationship between the first fusion result and the second fusion result, and correcting the third target load forecast result according to the nonlinear relationship to obtain the final load forecast result of the third level at the current moment, the expression of the final load forecast result is: In the formula, is the first fusion result, is the second fusion result, and W is the parameter of the neural network model.

[0074] In summary, the method of the present application is based on the historical time series collected by the central control platform and the valid data after exogenous variable expansion as input to build a multi-data input fusion short-term forecasting model that combines traditional machine learning and deep learning neural networks. It can accurately predict the load in different forecast time periods and realize the planning of demand, and use power load forecasts from different perspectives as features to improve the representation of short-term power load. Specifically, load data at the minute and hour levels are used, and they are combined with load data at the day level to improve the prediction accuracy of load values.

[0075] The absolute value error distribution bar graph of the experimental model and the control model is shown in Figure 3 From the figure, we can find that under different prediction time lengths, the error distribution of the experimental model is shifted to the left relative to the control group model, that is, the probability of a larger absolute error is smaller. As the prediction time becomes longer, the prediction accuracy of both models decreases, which is in line with the general prediction law. The prediction difficulty will increase as the prediction time becomes longer. Analyzing the situation when a larger error occurs, in the 1-day prediction scenario, the error of the experimental model is shifted to the left as a whole compared to the comparison model, indicating that the overall prediction accuracy of the experimental group is better than that of the control group. In addition, from the maximum value of the absolute error distribution, the maximum absolute error of the experimental group is about 2500, which is much smaller than the maximum error of the control group. Figure 3 As shown, it can be seen that in prediction scenarios of different durations, the experimental model performs significantly better than the control group model.

[0076] See also Figure 4 , which shows a structural block diagram of a power load forecasting system based on deep learning in the present application.

[0077] like Figure 4 As shown, the power load forecasting system 200 includes a filling module 210 , an acquisition module 220 , an input module 230 , a processing module 240 , an output module 250 , a fusion module 260 and a correction module 270 .

[0078] The filling module 210 is configured to obtain a first original data sequence uploaded by the central control platform within a first historical time period, and perform data filling on the first original data sequence using a preset missing value filling rule to obtain a first target original data sequence;

[0079] The acquisition module 220 is configured to acquire a second original data sequence reinforced input in a second historical time period, wherein the second historical time period does not overlap with the first historical time period, and the second historical time period is a historical time period adjacent to the current moment;

[0080] The input module 230 is configured to input the second original data sequence into the first RNN network layer of the pre-constructed power load forecasting model, and input the first target original data sequence into the first MLP network layer of the power load forecasting model, wherein the power load forecasting model is obtained through iterative training of the neural network, and the expression of the loss function of the iterative training of the neural network is:

[0081]

[0082] In the formula, is the penalty term, a is the weight coefficient, k is the number of predicted samples with negative values, w and b are self-optimizing parameters, n is the number of samples, and x k is the input vector when the predicted value is negative, and x is the input vector of the sample;

[0083] The processing module 240 is configured to perform external weight amplification on the output result of the first RNN network to obtain a first output result, and transmit the output result of the first MLP network layer and the first output result to the second MLP network layer of the power load forecasting model, and the second MLP network layer outputs a third output result;

[0084] The output module 250 is configured to input the third output result into at least one second RNN network layer of the power load forecasting model, and a second RNN network layer outputs a first target load forecast result of the first level at the current moment, a second target load forecast result of the second level at the current moment, and a third target load forecast result of the third level at the current moment;

[0085] A fusion module 260 is configured to fuse the first target load prediction result with the third target load prediction result based on a preset data fusion device to obtain a first fusion result, and to fuse the second target load prediction result with the third target load prediction result to obtain a second fusion result;

[0086] The correction module 270 is configured to use a preset neural network model to determine the nonlinear relationship between the first fusion result and the second fusion result, and to correct the third target load forecast result according to the nonlinear relationship to obtain a final load forecast result of the third level at the current moment. The expression of the final load forecast result is: In the formula, is the first fusion result, is the second fusion result, and w is the parameter of the neural network model.

[0087] It should be understood that Figure 4 Modules and references documented in Figure 1 Therefore, the operations and features described above for the method and the corresponding technical effects are also applicable to Figure 4 The modules in it will not be described in detail here.

[0088] In some other embodiments, the embodiments of the present invention further provide a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor executes the power load forecasting method based on deep learning in any of the above method embodiments;

[0089] As an implementation mode, the computer-readable storage medium of the present invention stores computer-executable instructions, and the computer-executable instructions are configured as follows:

[0090] Acquire a first original data sequence uploaded by the central control platform within a first historical time period, and perform data filling on the first original data sequence using a preset missing value filling rule to obtain a first target original data sequence;

[0091] Acquire a second original data sequence reinforced with input in a second historical time period, wherein the second historical time period does not overlap with the first historical time period, and the second historical time period is a historical time period adjacent to the current moment;

[0092] The second original data sequence is input into the first RNN network layer of the pre-built power load forecasting model, and the first target original data sequence is input into the first MLP network layer of the power load forecasting model, wherein the power load forecasting model is obtained through iterative training of the neural network, and the expression of the loss function of the iterative training of the neural network is:

[0093]

[0094] In the formula, is the penalty term, a is the weight coefficient, k is the number of predicted samples with negative values, w and b are self-optimizing parameters, n is the number of samples, and xk is the input vector when the predicted value is negative, and x is the input vector of the sample;

[0095] Performing external weight amplification on the output result of the first RNN network to obtain a first output result, and transmitting the output result of the first MLP network layer and the first output result to the second MLP network layer of the power load forecasting model, and the second MLP network layer outputs a third output result;

[0096] The third output result is input into at least one second RNN network layer of the power load forecasting model, and a second RNN network layer outputs a first target load forecasting result of the first level at the current moment, a second target load forecasting result of the second level at the current moment, and a third target load forecasting result of the third level at the current moment;

[0097] Based on a preset data fusion device, the first target load prediction result is fused with the third target load prediction result to obtain a first fusion result, and the second target load prediction result is fused with the third target load prediction result to obtain a second fusion result;

[0098] A preset neural network model is used to determine the nonlinear relationship between the first fusion result and the second fusion result, and the third target load forecast result is corrected according to the nonlinear relationship to obtain the final load forecast result of the third level at the current moment. The expression of the final load forecast result is: In the formula, is the first fusion result, is the second fusion result, and W is the parameter of the neural network model.

[0099] The computer-readable storage medium may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the power load forecasting system based on deep learning, etc. In addition, the computer-readable storage medium may include a high-speed random access memory, and may also include a memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the power load forecasting system based on deep learning via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0100] Figure 5 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 5As shown, the device includes: a processor 310 and a memory 320. The electronic device may also include: an input device 330 and an output device 340. The processor 310, the memory 320, the input device 330 and the output device 340 may be connected via a bus or other means. Figure 5 The example of the connection via bus is taken. The memory 320 is the above-mentioned computer-readable storage medium. The processor 310 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 320, that is, the power load forecasting method based on deep learning in the above-mentioned method embodiment is implemented. The input device 330 can receive input digital or character information, and generate key signal input related to user settings and function control of the power load forecasting system based on deep learning. The output device 340 may include display devices such as display screens.

[0101] The electronic device can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0102] As an implementation mode, the electronic device is applied to a power load forecasting system based on deep learning, and is used for a client, and includes: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can:

[0103] Acquire a first original data sequence uploaded by the central control platform within a first historical time period, and perform data filling on the first original data sequence using a preset missing value filling rule to obtain a first target original data sequence;

[0104] Acquire a second original data sequence reinforced with input in a second historical time period, wherein the second historical time period does not overlap with the first historical time period, and the second historical time period is a historical time period adjacent to the current moment;

[0105] The second original data sequence is input into the first RNN network layer of the pre-built power load forecasting model, and the first target original data sequence is input into the first MLP network layer of the power load forecasting model, wherein the power load forecasting model is obtained through iterative training of the neural network, and the expression of the loss function of the iterative training of the neural network is:

[0106]

[0107] In the formula, is the penalty term, a is the weight coefficient, k is the number of predicted samples with negative values, w and b are self-optimizing parameters, n is the number of samples, and x k is the input vector when the predicted value is negative, and x is the input vector of the sample;

[0108] Performing external weight amplification on the output result of the first RNN network to obtain a first output result, and transmitting the output result of the first MLP network layer and the first output result to the second MLP network layer of the power load forecasting model, and the second MLP network layer outputs a third output result;

[0109] The third output result is input into at least one second RNN network layer of the power load forecasting model, and a second RNN network layer outputs a first target load forecasting result of the first level at the current moment, a second target load forecasting result of the second level at the current moment, and a third target load forecasting result of the third level at the current moment;

[0110] Based on a preset data fusion device, the first target load prediction result is fused with the third target load prediction result to obtain a first fusion result, and the second target load prediction result is fused with the third target load prediction result to obtain a second fusion result;

[0111] A preset neural network model is used to determine the nonlinear relationship between the first fusion result and the second fusion result, and the third target load forecast result is corrected according to the nonlinear relationship to obtain the final load forecast result of the third level at the current moment. The expression of the final load forecast result is: In the formula, is the first fusion result, is the second fusion result, and W is the parameter of the neural network model.

[0112] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power load forecasting method based on deep learning, characterized in that: include: Acquire a first original data sequence uploaded by the central control platform within a first historical time period, and perform data filling on the first original data sequence using a preset missing value filling rule to obtain a first target original data sequence; Acquire a second original data sequence reinforced with input in a second historical time period, wherein the second historical time period does not overlap with the first historical time period, and the second historical time period is a historical time period adjacent to the current moment; The second original data sequence is input into the first RNN network layer of the pre-built power load forecasting model, and the first target original data sequence is input into the first MLP network layer of the power load forecasting model, wherein the power load forecasting model is obtained through iterative training of the neural network, and the expression of the loss function of the iterative training of the neural network is: In the formula, is the penalty term, a is the weight coefficient, k is the number of predicted samples with negative values, w and b are self-optimizing parameters, n is the number of samples, and x k is the input vector when the predicted value is negative, and x is the input vector of the sample; Performing external weight amplification on the output result of the first RNN network to obtain a first output result, and transmitting the output result of the first MLP network layer and the first output result to the second MLP network layer of the power load forecasting model, and the second MLP network layer outputs a third output result; The third output result is input into at least one second RNN network layer of the power load forecasting model, and a second RNN network layer outputs a first target load forecasting result of the first level at the current moment, a second target load forecasting result of the second level at the current moment, and a third target load forecasting result of the third level at the current moment; Based on a preset data fusion device, the first target load prediction result is fused with the third target load prediction result to obtain a first fusion result, and the second target load prediction result is fused with the third target load prediction result to obtain a second fusion result; A preset neural network model is used to determine the nonlinear relationship between the first fusion result and the second fusion result, and the third target load forecast result is corrected according to the nonlinear relationship to obtain the final load forecast result of the third level at the current moment. The expression of the final load forecast result is: In the formula, is the first fusion result, is the second fusion result, and W is the parameter of the neural network model.

2. The power load forecasting method based on deep learning according to claim 1, characterized in that: in, The first level, the second level and the third level are minute level, hour level and day level respectively.

3. The power load forecasting method based on deep learning according to claim 1, characterized in that: The expression of the first fusion result is: In the formula, is the first fusion result, is the first target load forecast result, is the power load characteristic of the third level at the current moment, is the latitude of the feature, is the third target load prediction result; The expression of the second fusion result is: In the formula, is the second fusion result, This is the second target load prediction result.

4. The power load forecasting method based on deep learning according to claim 1, characterized in that: The step of performing data filling on the first original data sequence by using a preset missing value filling rule to obtain a first target original data sequence includes: The initial first original data sequence is obtained, and the missing attributes are interpolated multiple times. In each iteration, the random forest is used to predict the value close to the missing value to replace it, and finally the complete first original data sequence is obtained.

5. The method for power load forecasting based on deep learning according to claim 1, characterized in that: Before inputting the second original data sequence into the first RNN network layer of the pre-built power load forecasting model, and inputting the first target original data sequence into the first MLP network layer of the power load forecasting model, the method further includes: Acquire an original data sequence within a preset historical time period, and perform data filling on the original data sequence using a preset missing value filling rule to obtain a target original data sequence, wherein the target original data sequence includes a target original data training subsequence and a target original data test subsequence; The target original data training subsequence is input into a neural network for iterative training to obtain an initial power load forecasting model, and the initial power load forecasting model is tested based on the target original data test subsequence to obtain a final power load forecasting model, wherein the power load forecasting model includes an inverse K-type network structure, and the inverse K-type network structure includes a first RNN network layer, a first MLP network layer, a second MLP network layer and at least one second RNN network layer.

6. The method for power load forecasting based on deep learning according to claim 1, characterized in that: The basic building block of the neural network is a neuron. The input x received by each neuron is multiplied by a weight w, then summed and added with a bias value b. An activation function f is applied to the sum to obtain the final output y. The expression of the output y is: y=f(∑wx+b), The activation function expression of the hidden layer in the neural network is: Where f(z) is the tanh function; The activation function expression of the output layer in the neural network is: Where f(z)′ is the sigmoid function.

7. The method for power load forecasting based on deep learning according to claim 1, characterized in that: Before inputting the second original data sequence into the first RNN network layer of the pre-built power load forecasting model, and inputting the first target original data sequence into the first MLP network layer of the power load forecasting model, the method further includes: The second original data sequence and the first target original data sequence are respectively normalized.

8. A power load forecasting system based on deep learning, characterized in that: include: a filling module, configured to obtain a first original data sequence uploaded by the central control platform within a first historical time period, and perform data filling on the first original data sequence using a preset missing value filling rule to obtain a first target original data sequence; an acquisition module, configured to acquire a second original data sequence reinforced by input in a second historical time period, wherein the second historical time period does not overlap with the first historical time period, and the second historical time period is a historical time period adjacent to the current moment; The input module is configured to input the second original data sequence into the first RNN network layer of the pre-constructed power load forecasting model, and input the first target original data sequence into the first MLP network layer of the power load forecasting model, wherein the power load forecasting model is obtained through iterative training of the neural network, and the expression of the loss function of the iterative training of the neural network is: In the formula, is the penalty term, a is the weight coefficient, k is the number of predicted samples with negative values, w and b are self-optimizing parameters, n is the number of samples, and x k is the input vector when the predicted value is negative, and x is the input vector of the sample; A processing module is configured to perform external weight amplification on the output result of the first RNN network to obtain a first output result, and transmit the output result of the first MLP network layer and the first output result to the second MLP network layer of the power load forecasting model, and the second MLP network layer outputs a third output result; An output module is configured to input the third output result into at least one second RNN network layer of the power load forecasting model, and a second RNN network layer outputs a first target load forecast result of the first level at the current moment, a second target load forecast result of the second level at the current moment, and a third target load forecast result of the third level at the current moment; a fusion module, configured to fuse the first target load prediction result with the third target load prediction result based on a preset data fusion device to obtain a first fusion result, and to fuse the second target load prediction result with the third target load prediction result to obtain a second fusion result; The correction module is configured to use a preset neural network model to determine the nonlinear relationship between the first fusion result and the second fusion result, and to correct the third target load forecast result according to the nonlinear relationship to obtain a final load forecast result of the third level at the current moment, wherein the expression of the final load forecast result is: In the formula, is the first fusion result, is the second fusion result, and W is the parameter of the neural network model.

9. The power load forecasting system based on deep learning according to claim 8, characterized in that: The system further comprises: The acquisition module is configured to obtain an original data sequence within a preset historical time period, and perform data filling on the original data sequence using a preset missing value filling rule to obtain a target original data sequence, wherein the target original data sequence includes a target original data training subsequence and a target original data test subsequence; The training module is configured to input the target original data training subsequence into the neural network for iterative training to obtain an initial power load forecasting model, and test the initial power load forecasting model based on the target original data test subsequence to obtain a final power load forecasting model, wherein the power load forecasting model includes an inverse K-type network structure, and the inverse K-type network structure includes a first RNN network layer, a first MLP network layer, a second MLP network layer and at least one second RNN network layer.

10. The power load forecasting system based on deep learning according to claim 9, characterized in that: The system further comprises: The normalization processing module is configured to perform normalization processing on the second original data sequence and the first target original data sequence respectively.

Images