Short-term power load prediction method considering data quality problem

Abstract

The invention discloses a short-term power load prediction method considering a data quality problem, and the method comprises the steps: firstly, providing an enhanced denoising auto-encoder based on bidirectional long and short-term memory, so as to extract time sequence characteristics which are robust to bad data in power load influence factor data; and then, taking the extracted features as the input of a time convolutional neural network, and establishing an end-to-end combined model to realize accurate load prediction. And finally, actual power load data is utilized to verify that the method still has relatively high prediction precision and applicability in a scene with a data quality problem. Compared with other short-term power load prediction methods, the short-term power load prediction method considering the data quality problem has the advantages that power load data containing the data quality problem can be rapidly and accurately predicted, and the predictability of dynamic behaviors of a power system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a short-term power load forecasting method, and in particular to a short-term power load forecasting method that takes data quality issues into consideration. Background Art

[0002] With the large-scale grid connection of renewable energy and the widespread application of large-capacity power electronic devices, the dynamic behavior of the power system has become more complex and changeable. The strong randomness of load and new energy has brought unprecedented challenges to the system regulation capability. Accurate short-term power load forecasting is of great significance to maintaining power balance, formulating reasonable dispatching plans and ensuring safe and stable operation of the power grid.

[0003] At present, the problem of short-term power load forecasting has been widely studied. The existing short-term power load forecasting methods are mainly divided into traditional statistical methods and artificial intelligence methods.

[0004] Statistical methods mainly include multivariate linear regression, Kalman filtering, and exponential smoothing. The goal of statistical methods is to establish a mathematical model for load forecasting based on the change pattern of current and historical load data through time series correlation. However, with the large-scale access of renewable energy, the randomness, intermittency and uncertainty of the system are gradually increasing, and traditional statistical methods are difficult to achieve accurate prediction of power load in complex nonlinear systems. In recent years, with the influx of a large amount of power big data, the advantages of short-term load forecasting methods based on artificial intelligence have gradually emerged. Existing artificial intelligence methods are mainly divided into machine learning methods and deep learning methods.

[0005] However, all of the above studies have considered the issue of data quality. When bad data exists in the power load data due to external factors such as communication or hardware equipment failure, the mapping between historical data and future power loads constructed by the existing short-term load forecasting method will be offset, and the error will increase sharply. The prediction results obtained will no longer have effective guiding significance and may even affect the safe and stable operation of the system.

[0006] Therefore, there is an urgent need for a short-term power load forecasting method that is effective for bad data in order to improve the scenario practicability of the algorithm and improve the prediction accuracy of power load. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the prior art and provide a short-term power load forecasting method that takes data quality issues into consideration. Compared with other short-term power load forecasting methods, the present invention can quickly and accurately predict power load data containing data quality issues, thereby improving the predictability of the dynamic behavior of the power system.

[0008] The technical solution to achieve the above object is: a short-term power load forecasting method considering data quality issues, comprising the following steps:

[0009] S1: Collect historical power load data and key influencing factors;

[0010] S2: Perform data normalization preprocessing on historical power load and key influencing factors;

[0011] S3: Construct an enhanced denoising autoencoder biLSTM-DAE and train it using the preprocessed data;

[0012] S4: Extracting features robust to bad data from power load data using the trained enhanced denoising autoencoder biLSTM-DAE;

[0013] S5: The obtained robust features are used as TCN input, and the power load forecast value at the next moment is obtained by further mapping it.

[0014] In the above-mentioned short-term power load forecasting method considering data quality issues, in step S1, key influencing factors include dry-bulb temperature, wet-bulb temperature, dew point temperature, humidity and electricity price.

[0015] In the above-mentioned short-term power load forecasting method considering data quality issues, in step S1, the time interval for collecting historical power load data is 30 minutes.

[0016] In the above-mentioned short-term power load forecasting method considering data quality issues, in step S2, the process of data normalization preprocessing is: preprocessing operations are performed on the power load data and key influencing factors:

[0017]

[0018] In formula (1), Z is the power load and different influencing factors data, Z max and Z min are the maximum and minimum values ​​of the corresponding variables, respectively.

[0019] In the above-mentioned short-term power load forecasting method considering data quality issues, in step S3, the enhanced denoising autoencoder biLSTM-DAE is a denoising autoencoder network constructed by a bidirectional LSTM.

[0020] In the above-mentioned short-term power load forecasting method considering data quality issues, in step S3, the process of constructing the enhanced denoising autoencoder biLSTM-DAE includes the following sub-steps:

[0021] S31: Let the normalized power load and key influencing factor data be In order to obtain historical power load and key influencing factor data with data quality problems, normal data are firstly Destruction to obtain historical power load and key influencing factor data with data quality issues Then, the original normal data x and the corrupted data Perform data preprocessing to obtain normalized and

[0022]

[0023] S32: The proposed bidirectional biLSTM enhanced denoising autoencoder is based on the normalized As input, it is encoded by a bidirectional biLSTM encoder to obtain a feature h that is sufficiently robust to bad data; then, another bidirectional biLSTM is used to reconstruct the feature h along the bidirectional time scale to obtain its reconstructed data The calculation process of the proposed enhanced denoising encoder is as follows:

[0024]

[0025] In formulas (2) and (3), biLSTM e and biLSTM d The operation processes of W are feature encoding and decoding reconstruction; e and W d They are biLSTM e and biLSTM d The weight matrix of e and b d They are biLSTM e and biLSTM d The bias vector of e and θ d The set of parameters of the entire network for encoding and decoding budgets;

[0026] S33: The proposed enhanced denoising autoencoder biLSTM-DAE is trained with the goal of minimizing the reconstruction error between the original normal historical power load and key influencing factor data and their corresponding reconstructed data. Its objective function is as follows:

[0027]

[0028] In formula (4), M is the batch size in the training process; N is the number of data points in the historical power load and key influencing factor data; NMSE is the normalized mean square error.

[0029] The above-mentioned short-term power load forecasting method takes into account data quality issues, wherein the smaller the reconstruction error, the stronger the ability of the extracted features to characterize the original normal data, and can provide a more effective feature basis for subsequent short-term load forecasting.

[0030] In the above-mentioned short-term power load forecasting method considering data quality issues, in step S4, the process of extracting features robust to bad data in power load data using the trained enhanced denoising autoencoder biLSTM-DAE is as follows: using the encoder biLSTM in the trained enhanced denoising autoencoder e Perform a high-dimensional mapping on the input data:

[0031]

[0032] In the formula, biLSTM e The operation process of W is feature encoding; e For biLSTM e The weight matrix of e For biLSTM e The bias vector of e It is the parameter set of the entire network during the encoding process.

[0033] In the above-mentioned short-term power load forecasting method considering data quality issues, in step S5, the robust time series features extracted are used as the input data of TCN. For each hidden layer, the extended causal convolution operation of the TCN with L layers at time t is defined as:

[0034]

[0035] In formula (5), represents the output of the lth layer TCN; W t (l) represents the TCN filter of the lth hidden layer; k represents the filter size; * represents the convolution operation; d represents the dilation factor; tj·d means that the data used for the convolution operation only depends on the input of the current time t and the past time.

[0036] The above-mentioned short-term power load forecasting method considering data quality issues, wherein the final forecast result is the output corresponding to the L-th layer TCN

[0037] Compared with other short-term power load forecasting methods, the short-term power load forecasting method considering data quality issues of the present invention can quickly and accurately predict power load data containing data quality issues, thereby improving the predictability of the dynamic behavior of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flow chart of a short-term power load forecasting method considering data quality issues of the present invention;

[0039] Figure 2 This is a structural diagram of a bidirectional biLSTM enhanced denoising autoencoder according to an embodiment of the present invention;

[0040] Figure 3 A structural diagram of a temporal convolutional neural network according to an embodiment of the present invention;

[0041] Figure 4 A schematic diagram showing a comparison of prediction results of different methods according to an embodiment of the present invention when the bad data ratio is 0%;

[0042] Figure 5 A schematic diagram showing a comparison of prediction results of different methods according to an embodiment of the present invention when the bad data ratio is 10%;

[0043] Figure 6 Schematic diagram showing comparison of prediction results of different methods according to an embodiment of the present invention when the bad data ratio is 15%;

[0044] Figure 7 It is a schematic diagram comparing prediction results of different methods according to an embodiment of the present invention when the bad data ratio is 20%. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the technical solution of the present invention, the specific implementation methods thereof are described in detail below in conjunction with the accompanying drawings:

[0046] See also Figures 1 to 3 , the best embodiment of the present invention, a short-term power load forecasting method considering data quality issues, comprises the following steps:

[0047] S1: Collect historical power load data and key influencing factors;

[0048] S2: Perform data normalization preprocessing on historical power load and key influencing factors;

[0049] S3: Construct an enhanced denoising autoencoder biLSTM-DAE and train it using the preprocessed data;

[0050] S4: Extracting features robust to bad data from power load data using the trained enhanced denoising autoencoder biLSTM-DAE;

[0051] S5: The obtained robust features are used as TCN input, and the power load forecast value at the next moment is obtained by further mapping it.

[0052] In step S1, the process of collecting historical power load data and key factors is as follows:

[0053] Firstly, the power load and key influencing factors data such as dry bulb temperature, wet bulb temperature, dew point temperature, humidity, and electricity price in a certain area from January 1, 2006 to December 31, 2010 for a total of 5 years were selected as experimental data, and the collection time interval of historical power load data was 30 minutes.

[0054] In step S2, the process of data normalization preprocessing is: preprocessing operations are performed on the power load data and key influencing factors:

[0055]

[0056] In formula (1), Z is the power load and different influencing factors data, Z max and Z min are the maximum and minimum values ​​of the corresponding variables, respectively.

[0057] In step S3, please refer to Figure 2 , the enhanced denoising autoencoder biLSTM-DAE is a denoising autoencoder network constructed by bidirectional LSTM. The process of constructing the enhanced denoising autoencoder biLSTM-DAE includes the following sub-steps:

[0058] S31: Let the normalized power load and key influencing factor data be In order to obtain historical power load and key influencing factor data with data quality problems, normal data are firstly Destruction to obtain historical power load and key influencing factor data with data quality issues Then, the original normal data x and the corrupted data Perform data preprocessing to obtain normalized and

[0059]

[0060] S32: The proposed bidirectional biLSTM enhanced denoising autoencoder is based on the normalized As input, it is encoded by a bidirectional biLSTM encoder to obtain a feature h that is sufficiently robust to bad data; then, another bidirectional biLSTM is used to reconstruct the feature h along the bidirectional time scale to obtain its reconstructed data The calculation process of the proposed enhanced denoising encoder is as follows:

[0061]

[0062] In formula (2) and formula (3), biLSTM e and biLSTM d The operation processes of W are feature encoding and decoding reconstruction; e and W d They are biLSTM e and biLSTM d The weight matrix of e and b d They are biLSTM e and biLSTM d The bias vector of e and θ d The set of parameters of the entire network for encoding and decoding budgets;

[0063] S33: The proposed enhanced denoising autoencoder biLSTM-DAE is trained with the goal of minimizing the reconstruction error between the original normal historical power load and key influencing factor data and their corresponding reconstructed data. Its objective function is as follows:

[0064]

[0065] In formula (4), M is the batch size in the training process; N is the number of data points in the historical power load and key influencing factor data; NMSE (Normalized Mean Square Error) is the normalized mean square error.

[0066] The smaller the reconstruction error, the stronger the ability of the extracted features to represent the original normal data, which can provide a more effective feature basis for subsequent short-term load forecasting.

[0067] In step S4, the process of extracting the features of the power load data that are robust to bad data using the trained enhanced denoising autoencoder biLSTM-DAE is as follows: using the encoder biLSTM in the trained enhanced denoising autoencoder e Perform a high-dimensional mapping on the input data:

[0068]

[0069] In the formula, biLSTM e The operation process of W is feature encoding; e For biLSTM e The weight matrix of e For biLSTM e The bias vector of e It is the parameter set of the entire network during the encoding process.

[0070] In step S5, please refer to Figure 3The obtained robust features are used as the input of the time convolutional neural network TCN. The specific process of obtaining the power load forecast value at the next moment by further mapping it is as follows: Based on the extracted robust time series features as the input data of TCN, the network structure diagram of TCN is as follows: Figure 3 As shown. For each hidden layer, the extended causal convolution operation at time t of a TCN with L layers is defined as:

[0071]

[0072] In formula (5), represents the output of the lth layer TCN; W t (l) represents the TCN filter of the lth hidden layer; k represents the filter size; * represents the convolution operation; d represents the dilation factor; tj·d means that the data used for the convolution operation only depends on the input of the current time t and the past time.

[0073] The final prediction result is the output corresponding to the Lth layer TCN

[0074] In order to illustrate the effect of the solution of the embodiment of the present invention, an experiment was carried out to verify it.

[0075] 1. Comparison of reconstruction error results of different autoencoders

[0076] In order to verify that the proposed enhanced denoising encoder has strong data reconstruction capability, different encoders are tested using normal power load and key influencing factor data. The normalized reconstruction errors of different encoders for various variables are shown in Table 1.

[0077] The normalized reconstruction error is expressed as follows:

[0078]

[0079] in, is the i-th preprocessed power load and key influencing factor data; for The reconstructed data is , N is the number of samples.

[0080] Table 1. Reconstruction error results of different autoencoders (10 -6 ):

[0081]

[0082] From Table 1 we can see that:

[0083] (1) From the reconstruction error of SAE and SDAE, it can be seen that the autoencoder with denoising structure can obtain lower reconstruction error, thereby improving the expressiveness of the encoding features;

[0084] (2) Comparing the reconstruction errors of SDAE, LSTM-DAE and the proposed method, it can be seen that the denoising autoencoder constructed by a neural network with a recursive structure can capture the long-term temporal dependency of different variables through the time dimension and effectively reduce the reconstruction error;

[0085] (3) By comparing the reconstruction errors of LSTM-DAE and the proposed method, it can be seen that the bidirectional biLSTM can further reduce the multivariate reconstruction error and improve the robustness of the feature to bad data by combining the bidirectional time scale to encode it.

[0086] 2. Short-term forecast results of actual power load

[0087] See also Figures 4 to 7 ,In order to verify that the proposed method still has good ,prediction accuracy in scenarios with data quality problems, the last day of power ,load is taken as an example, and the prediction results of the proposed ,method are compared with the traditional artificial intelligence algorithm and the ,existing short-term load forecasting methods under different proportions of bad ,data, where CNN-LSTM and CNN-biLSTM are the two existing hybrid load ,forecasting models. Figures 4 to 7 These are the power load forecast curves under different bad data ratios.

[0088] It can be seen that the prediction accuracy of the existing method is low at locations where the load fluctuation is more drastic, while the proposed method can still achieve load prediction with a lower prediction error. At the same time, as the proportion of bad data increases, the prediction accuracy of the existing method decreases sharply, seriously affecting the credibility of the algorithm. However, the proposed method can extract effective features with robustness from historical data with data quality problems, so it still has a high prediction accuracy when containing 20% ​​bad data.

[0089] 3. Comparison of actual power load short-term forecast results

[0090] In order to further show that the proposed method has higher robustness to data quality issues, the prediction accuracy of the proposed method is compared with other short-term load forecasting methods under different bad data ratios. Tables 2 and 3 are the RMSE and MAPE of each method under different bad data ratios, respectively, where the indicator deviation reflects the degree to which the method is affected by data quality issues.

[0091] Table 2, RMSE (kW) of different methods in scenarios with different data quality issues:

[0092]

[0093] Table 3, MAPE (%) of different methods in scenarios with different data quality issues:

[0094]

[0095]

[0096] It can be seen from Tables 2 and 3 that deep learning can better establish complex nonlinear mapping relationships and obtain better prediction accuracy compared to traditional machine learning. At the same time, the prediction method based on the hybrid model can combine the advantages of different neural networks to further improve the prediction accuracy. From the prediction errors and indicator deviations of different methods, it can be seen that the feature extraction method based on the denoising autoencoder can provide more robust and reliable features for short-term power load forecasting, so that accurate load prediction can be achieved with lower prediction errors in scenarios with bad data.

[0097] In summary, the short-term power load forecasting method considering data quality issues of the present invention, compared with other short-term power load forecasting methods, can quickly and accurately predict power load data containing data quality issues, thereby improving the predictability of the dynamic behavior of the power system.

[0098] Those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. As long as they are within the spirit of the present invention, any changes or modifications to the above embodiments will fall within the scope of the claims of the present invention.

Claims

1. A short-term power load forecasting method considering data quality issues, characterized in that: The following steps are involved: S1: Collect historical power load data and key influencing factors; S2: Perform data normalization preprocessing on historical power load and key influencing factors; S3: Construct an enhanced denoising autoencoder and train it using the preprocessed data; S4: Extracting features from power load data that are robust to bad data using the trained enhanced denoising autoencoder; S5: The obtained robust features are used as TCN input, and the power load forecast value at the next moment is obtained by further mapping it.

2. A short-term power load forecasting method considering data quality issues according to claim 1, characterized in that: In step S1, key influencing factors include dry bulb temperature, wet bulb temperature, dew point temperature, humidity and electricity price.

3. A short-term power load forecasting method considering data quality issues according to claim 1, characterized in that: In step S1, the historical power load data is collected every 30 minutes.

4. A short-term power load forecasting method considering data quality issues according to claim 1, characterized in that: In step S2, the process of data normalization preprocessing is: preprocessing operations are performed on the power load data and key influencing factors: In formula (1), Z is the power load and different influencing factors data, Z max and Z min are the maximum and minimum values ​​of the corresponding variables, respectively.

5. A short-term power load forecasting method considering data quality issues according to claim 1, characterized in that: In step S3, the enhanced denoising autoencoder is a denoising autoencoder network constructed by a bidirectional LSTM.

6. A short-term power load forecasting method considering data quality issues according to claim 1 or 5, characterized in that: In step S3, the process of constructing the enhanced denoising autoencoder includes the following sub-steps: S31: Let the normalized power load and key influencing factor data be In order to obtain historical power load and key influencing factor data with data quality problems, normal data are firstly Destruction to obtain historical power load and key influencing factor data with data quality issues Then, the original normal data x and the corrupted data Perform data preprocessing to obtain normalized and S32: The proposed bidirectional biLSTM enhanced denoising autoencoder is based on the normalized As input, it is encoded by a bidirectional biLSTM encoder to obtain a feature h that is sufficiently robust to bad data; then, another bidirectional biLSTM is used to reconstruct the feature h along the bidirectional time scale to obtain its reconstructed data The calculation process of the proposed enhanced denoising encoder is as follows: In formulas (2) and (3), biLSTM e and biLSTM d The operation processes of W are feature encoding and decoding reconstruction; e and W d They are biLSTM e and biLSTM d The weight matrix of e and b d They are biLSTM e and biLSTM d The bias vector of e and θ d The set of parameters of the entire network for encoding and decoding budgets; S33: The proposed enhanced denoising autoencoder is trained with the goal of minimizing the reconstruction error between the original normal historical power load and key influencing factor data and their corresponding reconstructed data. Its objective function is as follows: In formula (4), M is the batch size in the training process; N is the number of data points in the historical power load and key influencing factor data; NMSE is the normalized mean square error.

7. A short-term power load forecasting method considering data quality issues according to claim 6, characterized in that: The smaller the reconstruction error, the stronger the ability of the extracted features to represent the original normal data, which can provide a more effective feature basis for subsequent short-term load forecasting.

8. A short-term power load forecasting method considering data quality issues according to claim 1, characterized in that: In step S4, the process of extracting the features of the power load data that are robust to bad data using the trained enhanced denoising autoencoder is as follows: using the encoder biLSTM in the trained enhanced denoising autoencoder e Perform a high-dimensional mapping on the input data: In the formula, biLSTM e The operation process is feature encoding; W e For biLSTM e The weight matrix of e For biLSTM e The bias vector of e It is the parameter set of the entire network during the encoding process.

9. A short-term power load forecasting method considering data quality issues according to claim 1, characterized in that: In step S5, the extracted robust time series features are used as the input data of TCN. For each hidden layer, the extended causal convolution operation of TCN with L layers at time t is defined as: In formula (5), Represents the output of the lth layer TCN; represents the TCN filter of the lth hidden layer; k represents the filter size; * represents the convolution operation; d represents the dilation factor; tj·d means that the data used for the convolution operation only depends on the input of the current time t and the past time.

10. A short-term power load forecasting method considering data quality issues according to claim 9, characterized in that: The final prediction result is the output corresponding to the Lth layer TCN

Images