A QRNN-based Baseline Load Prediction Method and Device for Distribution Substations

Abstract

The present invention discloses a method and device for predicting baseline load of distribution substations based on QRNN. The method includes constructing a baseline load prediction model based on a QRNN network, selecting a number of load influencing factors as input feature vectors of the baseline load prediction model, and using the load forecast value of the day to be predicted as the output feature vector of the baseline load prediction model; in the baseline load prediction model, data preprocessing is performed on the input feature vector and the obtained data set is divided into a training set, a validation set, and a test set; the training set, the validation set, and the test set are calculated based on a preset prediction effect evaluation index and a preset hyperparameter to obtain a prediction result. The model proposed in the present invention combines the advantages of recurrent neural networks and convolutional neural networks, and has the characteristics of high precision, high computational efficiency, and parallel acceleration optimization, meeting the demand scenario of baseline load prediction of distribution substations under real-time and massive data.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a QRNN-based distribution station area baseline load prediction method and device. Background Art

[0002] Baseline load is the load consumed by users when they are not implementing demand response. By comparing it with actual load, the extent of user load reduction can be determined. Accurate and efficient baseline load forecasts provide the basis for quantitatively assessing the extent of user load reduction during demand response. This is a prerequisite for demand response implementers to penalize or compensate participants, and serves as one of the foundations for post-demand response effectiveness evaluations.

[0003] Currently, baseline load forecasting models include date matching, time series, regression analysis, and artificial intelligence forecasting methods, among others. Artificial intelligence forecasting methods primarily include machine learning and deep learning methods. In recent years, deep learning methods, with their powerful mapping capabilities for high-dimensional, nonlinear, and complex systems, have been widely used in the field of user baseline load forecasting. Key models include recurrent neural networks, convolutional neural networks, and other deep learning models. Deep learning methods can learn abstract features from massive amounts of data layer by layer through multi-layer networks. By combining massive amounts of historical load data with influencing factors such as weather and the economy, they can achieve more accurate load forecasting. However, the construction and training of large-scale networks poses greater challenges to computing speed. Summary of the Invention

[0004] The present invention provides a QRNN-based distribution station area baseline load prediction method, which can effectively improve the accuracy, efficiency and real-time performance of baseline load prediction.

[0005] In order to solve the above technical problems, an embodiment of the present invention provides a method for predicting baseline load of a distribution station area based on QRNN, comprising the following steps:

[0006] Constructing a baseline load forecasting model based on a QRNN network, selecting several load influencing factors as input feature vectors of the baseline load forecasting model, and using the load forecast value of the day to be forecasted as the output feature vector of the baseline load forecasting model;

[0007] In the baseline load forecasting model, the input feature vector is preprocessed and the obtained data set is divided into a training set, a validation set and a test set. The training set, the validation set and the test set are calculated based on preset prediction effect evaluation indicators and preset hyperparameters to obtain a prediction result.

[0008] As a further improvement, the QRNN network of the baseline load forecasting model based on the QRNN network is specifically:

[0009] The convolution part of the QRNN network:

[0010]

[0011] f t =σ(W f *X t )

[0012] o t =σ(W o *X t )

[0013] Pooling part of the QRNN network:

[0014]

[0015] h t =o t ⊙tanh(c t )

[0016] in, Represents the result of the m-dimensional time series convolution, the symbol * represents the masked convolution along the time step dimension, and the symbol ⊙ represents the element-wise multiplication operation; is the convolution filter, k is the width of the convolution filter, It means that the input feature vector consists of dimensions, c t is the multiplication value of the input feature vector, h t The output value of the operation.

[0017] As a further improvement, the load influencing factors are selected as the input feature vectors of the baseline load forecasting model, specifically:

[0018] The historical load parameter, the average temperature parameter, the relative humidity parameter, the date type parameter and the time value parameter are selected as the input feature vector of the baseline load forecasting model.

[0019] As a further improvement, the input feature vector is preprocessed and the obtained data set is divided into a training set, a validation set and a test set, specifically:

[0020] When the original characteristic data x(i) input into the baseline load forecasting model is preprocessed, the original characteristic data x(i) is normalized so that the input is limited to the range of [0, 1]. The normalized input data x1(i) is:

[0021]

[0022] Where xmax 、x min are the maximum and minimum values of the original input data respectively;

[0023] The missing values and "NAN" abnormal values that do not exceed 15% of the original feature data volume are filled with the mean of the adjacent load data; if continuous abnormal data appears, continue to search towards both ends until a non-empty value is found;

[0024] The obtained dataset was divided into training set, validation set and test set in a ratio of 70%:15%:15%.

[0025] As a further improvement, the preset prediction effect evaluation indicators include:

[0026] Mean absolute percentage error:

[0027]

[0028] Root mean square error:

[0029]

[0030] Among them, n is the number of prediction points, y i d i are the actual value and predicted value of the load at prediction point i respectively.

[0031] The embodiment of the present invention further provides a QRNN-based distribution station area baseline load prediction device, comprising:

[0032] A model building module is used to construct a baseline load forecasting model based on a QRNN network, select several load influencing factors as input feature vectors of the baseline load forecasting model, and use the load forecast value of the to-be-forecasted day as the output feature vector of the baseline load forecasting model;

[0033] The model data processing module is used to perform data preprocessing on the input feature vector in the baseline load forecasting model and divide the obtained data set into a training set, a validation set and a test set, and calculate the training set, the validation set and the test set based on the preset prediction effect evaluation index and the preset hyperparameters to obtain the prediction result.

[0034] As a further improvement, the QRNN network is specifically:

[0035] The convolution part of the QRNN network:

[0036]

[0037] f t =σ(W f *X t)

[0038] o t =σ(W o *X t )

[0039] Pooling part of the QRNN network:

[0040]

[0041] h t =o t ⊙tanh(c t )

[0042] in, Represents an m-dimensional time series, the symbol * represents a masked convolution along the time step dimension, and the symbol ⊙ represents an element-wise multiplication operation; is the convolution filter, k is the width of the convolution filter, It means that the input feature vector consists of dimensions, c t is the multiplication value of the input eigenvector, h t The output value of the operation.

[0043] As a further improvement, the input feature vector is specifically:

[0044] Historical load parameters, average temperature parameters, relative humidity parameters, date type parameters, and time value parameters.

[0045] As a further improvement, the model data processing module is further used to:

[0046] When the original characteristic data x(i) input into the baseline load forecasting model is preprocessed, the original characteristic data x(i) is normalized so that the input is limited to the range of [0, 1]. The normalized input data x1(i) is:

[0047]

[0048] Where x max 、x min are the maximum and minimum values of the original input data respectively;

[0049] The missing values and "NAN" abnormal values that do not exceed 15% of the original feature data volume are filled with the mean of the adjacent load data; if continuous abnormal data appears, continue to search towards both ends until a non-empty value is found;

[0050] The obtained dataset was divided into training set, validation set and test set in a ratio of 70%:15%:15%.

[0051] As a further improvement, the preset prediction effect evaluation indicators include:

[0052] Mean absolute percentage error:

[0053]

[0054] Root mean square error:

[0055]

[0056] Among them, n is the number of prediction points, y i d i are the actual value and predicted value of the load at prediction point i respectively.

[0057] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0058] By fully utilizing the existing hardware to build a demand response baseline load forecast based on the Quasi-Recurrent Neural Network (QRNN), it combines the advantages of recurrent neural networks and convolutional neural networks. It has the characteristics of high precision, high computational efficiency, and parallel acceleration optimization, meeting the demand scenarios of baseline load forecasting in distribution station areas under real-time and massive data.

[0059] The QRNN-based distribution area baseline load forecasting model can fully utilize the memory characteristics of the QRNN network and combine it with the time series characteristics of the load to achieve high prediction accuracy, which has an accuracy advantage over other popular deep learning models.

[0060] To address the problem that LSTM long short-term memory neural networks are difficult to optimize in parallel, resulting in low modeling and prediction rates, the QRNN model proposed in this invention can achieve parallel load forecasting by moderately simplifying the network structure, effectively improving prediction efficiency, and can be used for real-time, fast, and large-scale load forecasting, which is beneficial to engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 1 is a flowchart of the steps of the QRNN-based distribution station area baseline load prediction method in an embodiment of the present invention;

[0062] Figure 2 Schematic diagram of a QRNN-based distribution station area baseline load prediction method in an embodiment of the present invention;

[0063] Figure 3 is a unit structure diagram of a QRNN network in an embodiment of the present invention;

[0064] Figure 4This is a graph showing the prediction results of the three network test sets of QRNN, LSTM, and SVR in the embodiment of the present invention;

[0065] Figure 5 This is a diagram of the training and prediction time of the three network test sets of QRNN, LSTM, and SVR in the embodiment of the present invention. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0067] See Figure 1 The embodiment of the present invention provides a method for predicting baseline load of a distribution station area based on QRNN, comprising the following steps:

[0068] S1. Construct a baseline load forecasting model based on a QRNN network, select several load influencing factors as input feature vectors of the baseline load forecasting model, and use the load forecast value of the day to be forecasted as the output feature vector of the baseline load forecasting model;

[0069] S2. In the baseline load forecasting model, the input feature vector is preprocessed and the obtained data set is divided into a training set, a validation set and a test set. The training set, the validation set and the test set are calculated based on the preset prediction effect evaluation index and the preset hyperparameters to obtain the prediction results.

[0070] In an embodiment of the present invention, a baseline load prediction model for a distribution station area is first established based on QRNN. Secondly, load influencing factors are selected to construct input and output feature vectors, and the original data is preprocessed to obtain a training set, a validation set, and a test set. Furthermore, appropriate hyperparameters are selected for the network to optimize network performance. Finally, the established model is applied to perform demand response baseline load prediction to obtain prediction results.

[0071] The following combination Figure 2 , the specific implementation is described as follows.

[0072] 1. Design of Baseline Load Forecasting Model for Distribution Area Based on QRNN

[0073] 1) QRNN network structure

[0074] The QRNN quasi-recurrent neural network is a type of recurrent neural network, optimized from the LSTM long short-term memory network and the convolutional neural network (CNN). The LSTM is an improved RNN. By limiting the memory parameter to the interval [0, 1], it prevents the exponential impact of memories from more distant moments on the output. This effectively solves the vanishing gradient problem that RNNs cannot handle, fully utilizes historical information, and has greater adaptability in time series data analysis. The calculation method for the LSTM basic unit can be expressed as:

[0075] f t =σ(W f x t +U f h t-1 +b f )

[0076] i t =σ(W i x t +U i h t-1 +b i )

[0077] o t =σ(W o x t +U o h t-1 +b o )

[0078]

[0079]

[0080] h t =o t ⊙tanh(c t )

[0081] Where W is the matrix of network connection weights, W f x t , W i x t , W o x t , W c x t It is a matrix-vector multiplication, and its computational complexity is proportional to the size of the matrix, and the computational cost is much greater than equal-unit computation.

[0082] The inventors observed that LSTM contains 8 complex matrix-vector multiplications, where x is the input vector, so Wx can be combined into block matrix multiplications and then pre-calculated through parallel processing. However, h t-1Depends on the state of the hidden layer at the previous moment, so Uh t-1 The inability to pre-calculate forces LSTM to strictly follow sequential computations. While GPUs are being used to massively accelerate neural network training, LSTM cannot be accelerated through parallel learning. This limits LSTM's performance in real-time, fast, and large-scale load forecasting, hindering its engineering applications.

[0083] QRNN combines the advantages of long short-term memory (LSTM) and convolutional neural networks (CNN), reducing the computational complexity of loop steps in LSTM networks. Compared to the original LSTM, QRNN can fully parallelize computations, just like CNN.

[0084] The comparison between QRNN network structure and typical LSTM and CNN network structures is shown in the figure below. Figure 3 As shown: The brown part represents the matrix multiplication or convolution, the blue part is the parameter-free function that can be calculated in parallel in the feature or channel dimension, and the continuous and identical blocks mean that they can be executed in parallel. Figure 3 In the figure, LSTM can be decomposed into black linear blocks and white element blocks, and the calculation of each time step is related to the calculation result of the previous unit, so it cannot be calculated in parallel.

[0085] QRNN connects two types of layers, similar to the convolutional layer and pooling layer of CNN. Both types of layers allow for fully parallel computation. The convolutional layer supports parallelization across mini-batches and spatial dimensions (i.e., sequence dimensions), while the pooling layer supports parallelization across mini-batches and feature dimensions. The QRNN network of the baseline load forecasting model based on the QRNN network in step S1 is specifically as follows:

[0086] The convolution part of the QRNN network, these convolution operations produce time series

[0087]

[0088] f t =σ(W f *X t )

[0089] o t =σ(W o *X t )

[0090] The pooling part of the QRNN network has similar operations to the LSTM unit. In particular, the QRNN only uses the forget gate, also known as dynamic average pooling.

[0091]

[0092] h t =o t ⊙tanh(c t )

[0093] in, Represents the result of the m-dimensional time series convolution, the symbol * represents the masked convolution along the time step dimension, and the symbol ⊙ represents the element-wise multiplication operation; is the convolution filter, k is the width of the convolution filter, It means that the input feature vector consists of dimensions, c t is the multiplication value of the input feature vector, h t The output value of the operation.

[0094] From the calculation formula of QRNN, we can see that a unit in QRNN only performs three vector multiplication operations. These operations only depend on the input sequence and do not depend on the output of the previous unit, such as h t-1 . When the input is known, these vector multiplication operations can be pre-computed over multiple time steps. Therefore, there is no need to load the memory-intensive weight matrix at each time step. In this approach, the cost of DRAM decreases as the number of time steps required increases. Therefore, these operations can be parallelized in a single matrix-matrix multiplication, as follows:

[0095]

[0096] Where U∈R L×3d is the calculated merge matrix, d is the size of hidden layer neurons, L = T-k+1 is the sequence length. When considering the number of sub-training sample sets (mini batch) B, U∈R L×B×3d Constitute a tensor. The QRNN network construction algorithm can be described as Algorithm 1:

[0097]

[0098]

[0099] 2. Construction of Baseline Load Forecasting Model

[0100] 2.1. Input and output feature vector selection

[0101] Baseline load refers to the actual load of a user if they do not respond during a demand response day. This can be predicted based on historical load information and combined with factors such as weather and the environment to construct a demand response baseline load prediction model. Therefore, historical load parameters, average temperature parameters, relative humidity parameters, date type parameters, and time value parameters are selected as the input feature vector of the baseline load prediction model, totaling five dimensions. The input and output variables of the QRNN network are selected as shown in Table 1. The output vector is the predicted load value for the day to be predicted.

[0102] Table 1 Input and output variable selection

[0103]

[0104] 2.2 Data Preprocessing

[0105] The data preprocessing of the input feature vector and dividing the obtained data set into a training set, a validation set and a test set is specifically as follows:

[0106] When the original characteristic data x(i) input into the baseline load forecasting model is preprocessed, the original characteristic data x(i) is normalized so that the input is limited to the range of [0, 1]. The normalized input data x1(i) is:

[0107]

[0108] Where x max 、x min are the maximum and minimum values of the original input data respectively;

[0109] The missing values and "NAN" abnormal values that do not exceed 15% of the original feature data volume are filled with the mean of the adjacent load data; if continuous abnormal data appears, continue to search towards both ends until a non-empty value is found;

[0110] The obtained dataset was divided into training set, validation set and test set in a ratio of 70%:15%:15%.

[0111] 2.3 Prediction Evaluation Indicators

[0112] The preset prediction effect evaluation indicators include:

[0113] Mean absolute percentage error (MAPE):

[0114]

[0115] Root mean square error (RMSE):

[0116]

[0117] Among them, n is the number of prediction points, y i d i are the actual value and predicted value of the load at prediction point i respectively.

[0118] The following is an explanation of the implementation and results of short-term load forecasting.

[0119] To verify the performance of the proposed QRNN-based distribution network baseline load forecasting model, a comparative experiment was conducted. The experimental data was derived from load data from a distribution network in a city in Guangdong Province from May 27, 2019, to May 27, 2020, with a data step size of 15 minutes. The data was also derived from meteorological data provided by the local meteorological bureau, including average, maximum, and minimum temperature data, with a data step size of 24 hours.

[0120] Assuming that the meteorological data remains unchanged within a day, the meteorological data is extended periodically. The entire dataset will be divided into training set, validation set, and test set in a ratio of 70%:15%:15%.

[0121] Finally, the model input data is the load of the previous day, date type, time value and weather factors consisting of average temperature, minimum temperature and maximum temperature of the forecast point, and the input sequence length is 96; the output data is the single-point forecast load, and the output length is 1.

[0122] The experimental hardware platform used was an ECS.gn6i-c4gl.Xlarge cloud computing platform equipped with a 2.5GHz Intel Xeon Platinum 8163 CPU processor and an NVIDIA T4 GPU. The QRNN quasi-recurrent neural network was implemented using the TensorFlow (GPU version) 1.4.0 deep learning framework, using CUDA 8.0 for GPU parallel optimization. LSTM and SVR were used as control groups, both of which were implemented using the TensorFlow architecture.

[0123] For QRNN, LSTM and SVR, the Bayesian optimization algorithm [1] is used to search for the optimal hyperparameters of the model. Among them, the L2 regularization parameter of LSTM and QRNN models is set to 5×10 -6 , the dropout rate of each layer is 0.2.

[0124] The performance indicators of the experimental comparison algorithms include prediction accuracy and program running time. The day-ahead load forecasting experiment was conducted from May 20 to May 26. The prediction results of the three network test sets of QRNN, LSTM, and SVR are as follows: Figure 4 , the prediction error is shown in Table 2, and the training and prediction time is shown in Figure 5 and shown in Table 3.

[0125] Table 2 Comparison of test set errors

[0126]

[0127]

[0128] Table 3 Average running time

[0129]

[0130] Experimental results show that the QRNN and LSTM models significantly outperform the traditional SVR model in baseline load forecasting. Compared to the LSTM, the proposed QRNN model achieves slightly improved accuracy, while maintaining similar overall forecasting performance. Overall, QRNN, as a fusion of LSTM and CNN, demonstrates excellent forecasting accuracy.

[0131] In terms of running speed, the SVR network performed the fastest, followed by the QRNN network. Although the LSTM network had higher prediction accuracy, its training time was significantly longer than the other two algorithms. The average speed improvement rate of QRNN compared to LSTM under different data volumes is shown in Table 5. The average speed improvement rate index s is defined as:

[0132]

[0133] Where, t d and They are the average running time (training / testing) of QRNN and LSTM networks when the amount of data is d.

[0134] Table 4 Average speed improvement of QRNN compared to LSTM

[0135]

[0136] Table 4 shows that the QRNN-based baseline load forecasting model takes only 23.588 seconds to train on a single set of data, approximately 89.92% faster than the LSTM-based model. While the gap in execution speed on the test set has narrowed, the QRNN-based model is still 75.34% faster than the LSTM-based model. The SVR method trains slightly faster than the QRNN method, but its testing speed is much slower than the other two deep recurrent neural networks, resulting in poor performance in real-time baseline load forecasting.

[0137] As the amount of data increases, the QRNN method's acceleration relative to the LSTM method gradually increases. This means the QRNN method can better utilize computing resources and is more effective when the data volume is large. In actual baseline load forecasting platforms, models are typically updated within a fixed timeframe. When refined baseline load forecasting and analysis are required, fine-tuning a large number of different forecasting models consumes a significant amount of computing resources, posing a significant challenge to the operational efficiency of deep learning models such as LSTM. Experimental results demonstrate that the QRNN-based baseline load forecasting model proposed in this paper has excellent prediction accuracy and computational efficiency. On the one hand, the QRNN model's RNN-like structure enables it to reflect the connections between load data in time series, thereby deeply exploring the temporal correlations of load series. On the other hand, the QRNN model utilizes the principles of convolutional neural networks to improve the network structure and, combined with parallel optimization computation, significantly improves prediction efficiency. This enhances the engineering practicality of deep recurrent neural networks and provides a feasible solution for real-time, rapid, and large-scale baseline load forecasting.

[0138] The embodiment of the present invention further provides a QRNN-based distribution station area baseline load prediction device, comprising:

[0139] A model building module is used to construct a baseline load forecasting model based on a QRNN network, select several load influencing factors as input feature vectors of the baseline load forecasting model, and use the load forecast value of the to-be-forecasted day as the output feature vector of the baseline load forecasting model;

[0140] The model data processing module is used to perform data preprocessing on the input feature vector in the baseline load forecasting model and divide the obtained data set into a training set, a validation set and a test set, and calculate the training set, the validation set and the test set based on the preset prediction effect evaluation index and the preset hyperparameters to obtain the prediction result.

[0141] As a further improvement, the QRNN network is specifically:

[0142] The convolution part of the QRNN network:

[0143]

[0144] f t =σ(W f *X t )

[0145] o t =σ(W o *X t )

[0146] Pooling part of the QRNN network:

[0147]

[0148] h t =o t ⊙tanh(c t )

[0149] in, Represents an m-dimensional time series, the symbol * represents a masked convolution along the time step dimension, and the symbol ⊙ represents an element-wise multiplication operation; is the convolution filter, k is the width of the convolution filter, It means that the input feature vector consists of dimensions, c t is the multiplication value of the input eigenvector, h t The output value of the operation.

[0150] As a further improvement, the input feature vector is specifically:

[0151] Historical load parameters, average temperature parameters, relative humidity parameters, date type parameters, and time value parameters.

[0152] As a further improvement, the model data processing module is further used to:

[0153] When the original characteristic data x(i) input into the baseline load forecasting model is preprocessed, the original characteristic data x(i) is normalized so that the input is limited to the range of [0, 1]. The normalized input data x1(i) is:

[0154]

[0155] Where x max 、X min are the maximum and minimum values of the original input data respectively;

[0156] The missing values and "NAN" abnormal values that do not exceed 15% of the original feature data volume are filled with the mean of the adjacent load data; if continuous abnormal data appears, continue to search towards both ends until a non-empty value is found;

[0157] The obtained dataset was divided into training set, validation set and test set in a ratio of 70%:15%:15%.

[0158] As a further improvement, the preset prediction effect evaluation indicators include:

[0159] Mean absolute percentage error:

[0160]

[0161] Root mean square error:

[0162]

[0163] Among them, n is the number of prediction points, y i d i are the actual value and predicted value of the load at prediction point i respectively.

[0164] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A QRNN-based distribution area baseline load prediction method, characterized in that: The following steps are involved: Constructing a baseline load forecasting model based on a QRNN network, selecting several load influencing factors as input feature vectors of the baseline load forecasting model, and using the load forecast value of the day to be forecasted as the output feature vector of the baseline load forecasting model; In the baseline load forecasting model, the input feature vector is preprocessed and the obtained data set is divided into a training set, a validation set, and a test set. The training set, the validation set, and the test set are calculated based on a preset prediction effect evaluation index and preset hyperparameters to obtain a prediction result. The QRNN network of the baseline load forecasting model based on the QRNN network is specifically: The convolution part of the QRNN network: f t =σ(W f *X t ) the t =σ(W o *X t ) Pooling part of the QRNN network: h t =o t ⊙tanh(c t ) in, f t ,o t Represents the result of the m-dimensional time series convolution, the symbol * represents the masked convolution along the time step dimension, and the symbol ⊙ represents the element-wise multiplication operation; is the convolution filter, k is the width of the convolution filter, It means that the input feature vector consists of dimensions, c t is the multiplication value of the input feature vector, h t is the output value of the operation; The selection of several load influencing factors as input feature vectors of the baseline load forecasting model is specifically: Selecting historical load parameters, average temperature parameters, relative humidity parameters, date type parameters, and time value parameters as input feature vectors of the baseline load forecasting model; The data preprocessing of the input feature vector and dividing the obtained data set into a training set, a validation set and a test set is specifically as follows: When the original characteristic data x(i) input into the baseline load forecasting model is preprocessed, the original characteristic data x(i) is normalized so that the input is limited to the range of [0,1]. The normalized input data x1(i) is: Where x max 、x min are the maximum and minimum values of the original input data respectively; The missing values and "NAN" abnormal values that do not exceed 15% of the original feature data volume are filled with the mean of the adjacent load data; if continuous abnormal data appears, continue to search towards both ends until a non-empty value is found; The obtained dataset is divided into training set, validation set and test set in a ratio of 70%:15%:15%; The preset prediction effect evaluation indicators include: Mean absolute percentage error: Root mean square error: Among them, n is the number of prediction points, y i d i are the actual value and predicted value of the load at prediction point i respectively.

2. A QRNN-based distribution area baseline load prediction device, characterized in that: include: A model building module is used to construct a baseline load forecasting model based on a QRNN network, select several load influencing factors as input feature vectors of the baseline load forecasting model, and use the load forecast value of the to-be-forecasted day as the output feature vector of the baseline load forecasting model; a model data processing module, configured to perform data preprocessing on the input feature vector in the baseline load forecasting model and divide the obtained data set into a training set, a validation set, and a test set, and perform calculations on the training set, the validation set, and the test set based on preset prediction effect evaluation indicators and preset hyperparameters to obtain a prediction result; The QRNN network is specifically: The convolution part of the QRNN network: f t =σ(W f *X t ) the t =σ(W o *X t ) Pooling part of the QRNN network: h t =o t ⊙tanh(c t ) in, f t ,o t Represents the result of the m-dimensional time series convolution, the symbol * represents the masked convolution along the time step dimension, and the symbol ⊙ represents the element-wise multiplication operation; is the convolution filter, k is the width of the convolution filter, It means that the input feature vector consists of dimensions, c t is the multiplication value of the input feature vector, h t is the output value of the operation; The input feature vector is specifically: Historical load parameters, average temperature parameters, relative humidity parameters, date type parameters, and time value parameters; Wherein, the model data processing module is further used for: When the original characteristic data x(i) input into the baseline load forecasting model is preprocessed, the original characteristic data x(i) is normalized so that the input is limited to the range of [0,1]. The normalized input data x1(i) is: Where x max 、x min are the maximum and minimum values of the original input data respectively; The missing values and "NAN" abnormal values that do not exceed 15% of the original feature data volume are filled with the mean of the adjacent load data; if continuous abnormal data appears, continue to search towards both ends until a non-empty value is found; The obtained dataset is divided into training set, validation set and test set in a ratio of 70%:15%:15%; The preset prediction effect evaluation indicators include: Mean absolute percentage error: Root mean square error: Among them, n is the number of prediction points, y i d i are the actual value and predicted value of the load at prediction point i respectively.

Images