A method and device for repairing power positive active total value based on NeuralProphet and a medium

Abstract

The application discloses a power positive active total value repairing method and device based on NeuralProphet, and a medium, the method comprises the following steps: acquiring power positive active total values at each moment as original data; preprocessing the original data, wherein the preprocessing comprises deleting abnormal power positive active total values; constructing a training set according to the original data after preprocessing; constructing a prediction network model with NeuralProphet as a skeleton network; training the prediction network model through the training set; predicting the deleted power positive active total values through the trained prediction network model; and repairing the original data through the predicted power positive active total values; the application can accurately predict the power positive active total values and repair the original data, thereby ensuring accurate metering of power data.

Description

TECHNICAL FIELD

[0001] The application relates to a NeuralProphet-based power positive active total value repairing method, device and medium, and belongs to the technical field of data processing. BACKGROUND

[0002] A new power system is a prerequisite for carrying out the new development concept, constructing a new development pattern, and promoting high-quality development, ensuring energy and power safety as a basic prerequisite, taking meeting the power demand of economic and social development as the primary goal, taking maximizing the consumption of new energy as the main task, taking a strong smart grid as the hub platform, taking source-grid-load-storage interaction and multi-energy complementation as support, and having the basic characteristics of clean and low-carbon, safe and controllable, flexible and efficient, intelligent and friendly, and open and interactive. In the construction of the new power system, the problem of changes in power load structure and characteristics is faced, and the future power demand continues to grow, so it is very important to accurately measure power data.

[0003] In actual power measurement, positive active total value is recorded, and positive indicates that the power grid supplies power to users, and active is the actual power used by users. The positive active total value represents the power transmitted by the power grid to the user. The power meter is a tool for measuring the positive active total value. In a meter reading cycle, the user's power consumption is determined by the change in the positive active total value. However, the power meter measurement may also have data anomalies, which may affect the construction of the new power system. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art, provide a NeuralProphet-based power positive active total value repairing method, device and medium, and solve the technical problem that if the power meter measurement has data anomalies, the power data cannot be accurately measured, thereby affecting the construction of the new power system.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] In a first aspect, the present application provides a NeuralProphet-based power positive active total value repairing method, comprising:

[0007] Obtaining power positive active total values at each time as original data;

[0008] Preprocessing the original data, the preprocessing including deleting abnormal power positive active total values;

[0009] Constructing a training set according to the preprocessed original data;

[0010] A prediction network model is constructed with NeuralProphet as a skeleton network;

[0011] The prediction network model is trained through the training set;

[0012] The deleted power forward active total value is predicted through the trained prediction network model;

[0013] The original data is repaired through the predicted power forward active total value.

[0014] Optionally, the abnormal power forward active total value includes a power forward active total value with an incomplete field or a null value.

[0015] Optionally, constructing the training set according to the preprocessed original data includes: taking each time point in the preprocessed original data as an input sample, and taking the power forward active total value at each time point as the true label of the input sample; and the input sample and the true label thereof are one-to-one corresponding to form the training set.

[0016] Optionally, training the prediction network model through the training set includes:

[0017] determining model parameters of the prediction network model, the model parameters including a learning rate of a trend module, a seasonal pattern of a seasonal module, and an order of an autoregressive module;

[0018] setting the seasonal pattern and the order to be constant, setting a learning rate range and a learning rate update gradient, calculating a prediction effect through the training set and the prediction network model for each learning rate, taking the best learning rate as η1, determining learning rates η2 and η3 in a preset range according to the learning rate η1: η2 = η1-Δη; η2 = η1+Δη, and Δη is a range variable;

[0019] setting the seasonal pattern and the learning rate to be constant, setting an order range and an order update gradient, calculating a prediction effect through the training set and the prediction network model for each order, taking the best order as P1, determining orders P2 and P3 in a preset multiple according to the order P1: P2 = (n-1)T, P3 = (n+1)T, n = P1 / T; T = t max -t min , t max , t min are the maximum time point and the minimum time point in the training set, respectively;

[0020] the addition mode and the multiplication mode in the seasonal pattern are denoted as m1 and m2, respectively, and one is extracted from the learning rates η1, η2, η3, the seasonal patterns m1, m2, and the orders P1, P2, P3, respectively, for permutation and combination;

[0021] All combinations are obtained, for each combination, a prediction effect is calculated by the training set and the prediction network model, and the combination with the best prediction effect is taken as the final model parameter;

[0022] The final model parameter is brought into the prediction network model to obtain the trained prediction network model.

[0023] Optionally, the calculation of the prediction effect by the training set and the prediction network model comprises:

[0024] The input sample in the training set is input into the prediction network model to obtain a predicted value;

[0025] The mean absolute error (MAE) of the real label and the predicted value of the input sample is calculated:

[0026]

[0027] In the formula, y i 、 The real label and the predicted value of the i-th input sample, respectively, and I is the number of input samples in the training set;

[0028] The prediction effect is measured by the mean absolute error (MAE), and the smaller the mean absolute error (MAE), the better the prediction effect.

[0029] In a second aspect, the present application provides a power positive active total value repairing device based on NeuralProphet, the device comprises:

[0030] A data acquisition module is configured to acquire power positive active total values at each time as original data;

[0031] A preprocessing module is configured to preprocess the original data, and the preprocessing comprises deleting abnormal power positive active total values;

[0032] A training set module is configured to construct a training set according to the preprocessed original data;

[0033] A model construction module is configured to construct a prediction network model with NeuralProphet as a skeleton network;

[0034] A model training module is configured to train the prediction network model by the training set;

[0035] A model prediction module is configured to predict the deleted power positive active total values by the trained prediction network model;

[0036] A data repairing module is configured to repair the original data by the predicted power positive active total values.

[0037] In a third aspect, the present application provides a device for repairing power positive active total value based on NeuralProphet, comprising a processor and a storage medium.

[0038] The storage medium is used to store instructions.

[0039] The processor is used to operate according to the instructions to perform the steps of the above method.

[0040] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the steps of the above method.

[0041] Compared with the prior art, the present application has the following beneficial effects:

[0042] The present application provides a NeuralProphet-based power positive active total value repair method, device and medium, which constructs a prediction network model with NeuralProphet as the skeleton network, wherein the trend module, the seasonal module and the autoregressive module each have a separate input and modeling process, and each contributes an additional component to the final prediction of the prediction network model for the power positive active total value, realizing effective prediction of the power positive active total value at a specific time. The predicted positive active total value can be used to modify the power positive active total value outliers and fill in the power positive active total value nulls, thereby solving the technical problem that if the data of the electric meter is abnormal, the power data cannot be accurately metered. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a flowchart of a NeuralProphet-based power positive active total value repair method provided by the present application embodiment one;

[0044] Figure 2 is a training process flowchart of the prediction network model provided by the present application embodiment one. DETAILED DESCRIPTION

[0045] The present application will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0046] Embodiment one:

[0047] As shown in Figure 1 , the present application provides a NeuralProphet-based power positive active total value repair method, comprising the following steps:

[0048] 1. Obtain the power positive active total value at each time as the original data.

[0049] 2. Preprocessing the original data, the preprocessing includes deleting abnormal power positive forward total value;

[0050] The abnormal power positive forward total value includes but is not limited to the power positive forward total value with incomplete or null field.

[0051] 3. Constructing a training set according to the preprocessed original data;

[0052] Taking each time point in the preprocessed original data as an input sample, and taking the power positive forward total value at each time point as the true label of the input sample; the input sample and its true label are one-to-one corresponding to form a training set.

[0053] 4. Constructing a prediction network model with NeuralProphet as the skeleton network.

[0054] NeuralProphet includes a trend module, a seasonal module, and an autoregressive module.

[0055] (1) The trend module defines the trend changing with time. The positive forward total value data is a time series, which has a certain relationship with time t. The growth rate and the offset are combined to model the trend component. The trend is described by defining a growth rate δ(t) changing with time and a bias ρ(t) changing with time. The trend effect value T(t) of the positive forward total value time series in the trend module is calculated as follows:

[0056] T(t) = δ(t) · t + ρ(t)

[0057] The learning rate is set to fit the positive forward total value data, and the trend effect value T(t) of the positive forward total value to be predicted is calculated.

[0058] (2) The seasonal module defines the trend changing with season. The positive forward total value appears regular changes in different intervals of the same time period such as year, week, and day. The original seasonality in NeuralProphet is processed by using Fourier transform, but Fourier transform is not good enough for non-stationary signals. Therefore, the original seasonality module is processed by using wavelet transform. A wavelet transform function is defined for each season, and the seasonal effect value S(t) of each season is calculated as follows: p

[0059]

[0060] Where, a j，k , ψ j，k ​(t) is a wavelet transform term, which is fitted from the input forward aggregate active value data; p is a seasonal period, k is the number of transform terms using wavelet functions in the transform term for the season with seasonal period p. In multiple seasonal scenarios, different values can be defined for each period. The correspondence between p and k is shown in Table 1. Each wavelet function transform term corresponds to a combination to model.

[0061] Table 1:

[0062]

[0063]

[0064] NeuralProphet supports both additive and multiplicative seasonal patterns, with additive as the default, indicating that there is no heteroscedasticity in the predicted forward aggregate active value series. If there is obvious variance in the forward aggregate active value series, it can be configured as multiplicative. If configured as multiplicative, the component is multiplied by the trend effect, which is represented as follows:

[0065]

[0066] Daily, weekly, or yearly seasonality is automatically activated according to the data frequency and length. If the resolution of the obtained forward aggregate active value data frequency is higher than a certain periodicity, and there are at least two forward aggregate active value data of this periodicity, then this single seasonal period is activated. If the forward aggregate active value data is daily frequency, and there are two weeks or more of forward aggregate active value data, then the weekly seasonality is activated, and the daily seasonality is not activated, because the data frequency is not higher than the daily periodicity, and the annual seasonality is not activated, because there are not two periods of data. For time t, the impact of all seasonal factors considered in the seasonal module is calculated to obtain the seasonal effect value S(t) of the predicted forward aggregate active value:

[0067]

[0068] where P is the set of all periodicities, is the effect value of the activated single seasonal period.

[0069] (3) The autoregressive module defines the autoregressive effect according to past observations at time t, which is obtained by regressing the future value of the forward aggregate active value with the past value through autoregression (AR). A coefficient θ i is fitted for each past value to control the influence of a specific past value on the predicted direction and amplitude. The number of past values included is usually referred to as the order p of the AR model. In a classic autoregressive augmented reality process, an intercept c and a white noise term ε tThe autoregressive effect value A(t) is calculated considering the inside:

[0070]

[0071] If the traditional AR model is used to predict the autoregressive effect value A(t) of the positive total active value, only one-step prediction value can be obtained. If multi-step prediction of h>1 is needed, h models must be fitted, one model for each prediction step. The AR module in NeuralProphet is based on a modified version of AR-Net, which can generate multiple predictions with one model. The input is the positive total active value, and the output is the value of the AR effect of the positive total active value predicted from time t to time t+h-1 t (t), A t (t+1),..., A t (t+h-1):

[0072] A t (t), A t (t+1),..., A t (t+h-1) = AR-Net(y t-1 , y t-2 ,..., y t-p )

[0073] The model can be explained as:

[0074] y = Wx

[0075] where x, y are as follows:

[0076]

[0077] The positive total active value is taken as input x∈R p , the weight matrix W∈R h×p , W i，j represents the coefficient of the linear influence of y j on the positive total active value AR effect A t (t=i), which is obtained by fitting the input positive total active value, so as to calculate the value of the predicted positive total active value AR effect.

[0078] The most important parameter of this module is the number of past positive total active values to be regressed, also known as the order p of the AR model, which is usually set to twice the prediction period.

[0079] After the effect value of the corresponding module is obtained by predicting the positive total active value through the above module, the effect values are combined to obtain the predicted positive total active value

[0080] y t = T(t) + S(t) + A(t)

[0081] Wherein, T(t) is the predicted trend effect value in the trend module, S(t) is the seasonal effect value in the seasonal module, and A(t) is the autoregressive effect value in the autoregressive module.

[0082] When the power positive active total value is predicted by using NeuralProphet, the learning rate of the trend module, the seasonal pattern of the seasonal module, and the order of the autoregressive module can be set by themselves, and the prediction effects obtained by different values are different. The prediction effect can be better by selecting the parameters.

[0083] 5. Training the prediction network model by using the training set;

[0084] As shown in the following table: Figure 2 The training process is as follows:

[0085] 5.1. Determine the model parameters of the prediction network model, including the learning rate of the trend module, the seasonal pattern of the seasonal module, and the order of the autoregressive module.

[0086] 5.2. Set the seasonal pattern and order and keep them unchanged, set the learning rate range (η min , η max ) and the learning rate update gradient Δη', for each learning rate {η 1 , η 2 , …, η M}; η m+1 = η m + Δη'; m = 1, 2, …, M-1; η 1 = η min ; η M = η max , calculate the prediction effect by using the training set and the prediction network model, and record the best learning rate as η1; according to the learning rate η1, determine the learning rates η2 and η3 in the preset range: η2 = η1- Δη; η2 = η1+ Δη, and Δη is a range variable;

[0087] 5.3. Set the seasonal pattern and learning rate and keep them unchanged, set the order range (P min , P max ) and the order update gradient T, for each order {P 1 , P 2 , …, P S}; P s+1 = P s + T; s = 1, 2, …, S-1; P 1 = P min = T; η S = P max= ST, the best prediction effect of the order is recorded as P1 by calculating the prediction effect through the training set and the prediction network model; the orders P2 and P3 in the preset multiple are determined according to the order P1: P2 = (n-1)T, P3 = (n+1)T, n = P1 / T; T = t max -t min , t max , t min are the maximum time and the minimum time in the training set, respectively;

[0088] 5.4, the addition mode and the multiplication mode in the seasonal mode are recorded as m1 and m2, respectively, and one is extracted from the learning rate η1, η2, η3, the seasonal mode m1, m2, the order P1, P2, P3, respectively, for permutation combination;

[0089] 5.5, all combinations are obtained, for each combination, the prediction effect is calculated by the training set and the prediction network model, and the best combination of the prediction effect is taken as the final model parameter;

[0090] All combinations are shown in Table 2:

[0091] Table 2:

[0092]

[0093]

[0094] 5.6, the final model parameter is brought into the prediction network model to obtain the trained prediction network model.

[0095] In steps 5.2, 5.3 and 5.5, the prediction effect is calculated by the training set and the prediction network model, including:

[0096] The input sample in the training set is input into the prediction network model to obtain the predicted value;

[0097] The mean absolute error MAE of the real label and the predicted value of the input sample is calculated:

[0098]

[0099] In the formula, y i , are the real label and the predicted value of the i-th input sample, respectively, and I is the number of input samples in the training set;

[0100] The prediction effect is measured by the mean absolute error MAE, and the smaller the mean absolute error MAE, the better the prediction effect.

[0101] 6, the deleted power positive total value is predicted by the trained prediction network model.

[0102] 7. The original data is repaired by the predicted positive total value of the power.

[0103] Embodiment two:

[0104] The embodiment of the application provides a device for repairing a positive total value of power based on NeuralProphet, which comprises:

[0105] A data acquisition module is configured to acquire the positive total value of power at each time as original data.

[0106] A preprocessing module is configured to preprocess the original data, and the preprocessing comprises deleting abnormal positive total value of power.

[0107] A training set module is configured to construct a training set according to the preprocessed original data.

[0108] A model construction module is configured to construct a prediction network model with NeuralProphet as a skeleton network.

[0109] A model training module is configured to train the prediction network model through the training set.

[0110] A model prediction module is configured to predict the deleted positive total value of power through the trained prediction network model.

[0111] A data repairing module is configured to repair the original data through the predicted positive total value of power.

[0112] Embodiment three:

[0113] Based on embodiment one, the embodiment of the application provides a device for repairing a positive total value of power based on NeuralProphet, which comprises a processor and a storage medium.

[0114] The storage medium is configured to store instructions.

[0115] The processor is configured to operate according to the instructions to perform the steps of the above method.

[0116] Embodiment four:

[0117] Based on embodiment one, the embodiment of the application provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps of the above method.

[0118] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0119] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0120] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0122] The above description is only preferred embodiments of the application. It should be pointed out that, for those skilled in the art, some improvements and modifications can be made without departing from the technical principles of the application, and these improvements and modifications should also be considered as falling within the scope of the application.

Claims

1. A method for repairing the total positive active power value based on NeuralProphet, characterized in that, include: Obtain the total positive active power at each moment as the raw data; The raw data is preprocessed, including the removal of abnormal positive active power total values; A training set is constructed based on the preprocessed raw data; Construct a prediction network model with NeuralProphet as the backbone network; The prediction network model is trained using the training set; Predict the total positive active power to be removed using a trained prediction network model; The original data is repaired by predicting the total positive active power of electricity. The step of training the prediction network model using a training set includes: Determine the model parameters of the prediction network model, including the learning rate of the trend module, the seasonal pattern of the seasonal module, and the order of the autoregressive module; Set the seasonal mode and order and keep them unchanged. Set the learning rate range and learning rate update gradient. For each learning rate, calculate the prediction effect through the training set and the prediction network model. Record the learning rate with the best prediction effect as η1. Determine the learning rates η2 and η3 within the preset range based on the learning rate η1: η2 = η1 - Δη; η3 = η1 + Δη, where Δη is the range variable. Set the seasonal mode and learning rate and keep them constant. Define the order range and order update gradient. For each order, calculate the prediction performance using the training set and the prediction network model. Record the order with the best prediction performance as P1. Determine the orders P2 and P3 within the preset scaling factor based on order P1: P2 = (n-1)T, P3 = (n+1)T, n = P1 / T; T = t max -t min , t max t min These are the maximum and minimum times in the training set, respectively; Let the addition mode and multiplication mode in the seasonal pattern be denoted as m1 and m2 respectively, and draw one from the learning rate η1, η2, η3, seasonal pattern m1, m2, and order P1, P2, P3 respectively for permutation and combination. Obtain all combinations, calculate the prediction effect for each combination using the training set and the prediction network model, and use the combination with the best prediction effect as the final model parameters; The final model parameters are then fed into the prediction network model to obtain the trained prediction network model.

2. The method for repairing the total positive active power value based on NeuralProphet according to claim 1, characterized in that, The abnormal total positive active power value includes total positive active power values ​​with incomplete or null fields.

3. The method for repairing the total positive active power value based on NeuralProphet according to claim 1, characterized in that, The construction of the training set based on the preprocessed raw data includes: taking each time step in the preprocessed raw data as the input sample, and the total positive active power at each time step as the true label of the input sample; and constructing a training set by matching the input samples and their true labels one by one.

4. The method for repairing the total positive active power value based on NeuralProphet according to claim 1, characterized in that, The calculation of prediction results using the training set and the prediction network model includes: Input samples from the training set are fed into the prediction network model to obtain predicted values; Calculate the mean absolute error (MAE) between the true labels and predicted values ​​of the input samples: In the formula, y i , Let be the true label and the predicted value of the i-th input sample, respectively, and I be the number of input samples in the training set; The mean absolute error (MAE) is used to measure the prediction performance; the smaller the MAE, the better the prediction performance.

5. A device for repairing the total positive active power value based on NeuralProphet, characterized in that, The device includes: The data acquisition module is used to acquire the total positive active power value at each time point as raw data; The preprocessing module is used to preprocess the raw data, and the preprocessing includes deleting abnormal positive active power total values; The training set module is used to construct a training set based on the preprocessed raw data. The model building module is used to build prediction network models with NeuralProphet as the backbone network. The model training module, used to train the prediction network model using the training set, includes: Determine the model parameters of the prediction network model, including the learning rate of the trend module, the seasonal pattern of the seasonal module, and the order of the autoregressive module; Set the seasonal mode and order and keep them unchanged. Set the learning rate range and learning rate update gradient. For each learning rate, calculate the prediction effect through the training set and the prediction network model. Record the learning rate with the best prediction effect as η1. Determine the learning rates η2 and η3 within the preset range based on the learning rate η1: η2 = η1 - Δη; η3 = η1 + Δη, where Δη is the range variable. Set the seasonal mode and learning rate and keep them constant. Define the order range and order update gradient. For each order, calculate the prediction performance using the training set and the prediction network model. Record the order with the best prediction performance as P1. Determine the orders P2 and P3 within the preset scaling factor based on order P1: P2 = (n-1)T, P3 = (n+1)T, n = P1 / T; T = t max -t min , t max t min These are the maximum and minimum times in the training set, respectively; Let the addition mode and multiplication mode in the seasonal pattern be denoted as m1 and m2 respectively, and draw one from the learning rate η1, η2, η3, seasonal pattern m1, m2, and order P1, P2, P3 respectively for permutation and combination. Obtain all combinations, calculate the prediction effect for each combination using the training set and the prediction network model, and use the combination with the best prediction effect as the final model parameters; The final model parameters are then fed into the prediction network model to obtain the trained prediction network model. The model prediction module is used to predict the total positive active power to be removed using a trained prediction network model. The data repair module is used to repair the original data by using the predicted total positive active power value.

6. A device for repairing the total positive active power value based on NeuralProphet, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-4.

Images