A method and device for detecting electricity theft users in a low-voltage distribution network

Abstract

The present invention belongs to the technical field of encoders, and discloses a method and device for detecting electricity theft users in a low-voltage distribution network. The detection method comprises: obtaining user data to be detected and preprocessing the user data to be detected; extracting features from the preprocessed user data to be detected to obtain a feature data set of the user to be detected; inputting the feature data set of the user to be detected into a pre-trained Wide and Deep model for detection to obtain electricity theft users. Beneficial effect: After obtaining the user data to be detected in the area to be detected, preprocessing and feature extraction are performed and input into the preset Wide and Deep model, electricity theft users can be efficiently and accurately screened out, thereby ensuring the economic benefits of the power grid enterprise.

Description

Technical Field

[0001] The invention relates to the technical field of encoders, and in particular to a method and a device for detecting electricity theft users in a low-voltage distribution network. Background Art

[0002] In the daily operation of low-voltage distribution networks, line loss management will affect the accuracy of load demand forecasting of the power grid and thus lead to safety accidents in the operation of the power grid. Line loss mainly consists of two parts, namely technical line loss and non-technical line loss. Non-technical line loss refers to the uncalculated part of the power transmission and distribution loss of the power grid after deducting the reasonable technical line loss. The user's breach of contract, especially theft of electricity, is the main cause of non-technical line loss.

[0003] Since the release of Document No. 9 on the Electricity Reform in 2015, the economic benefits of power grid companies have been hit unprecedentedly as the power reform has continued. Against this background, it has become an urgent need for power grid companies to increase electricity sales revenue by reducing electricity theft.

[0004] Therefore, designing an efficient and accurate electricity theft detection method based on actual application scenarios is an urgent problem to be solved. Summary of the invention

[0005] The purpose of the present invention is to provide a method and device for detecting electricity theft users in a low-voltage distribution network according to actual application scenarios, so as to efficiently and accurately screen out electricity theft users and ensure the economic benefits of power grid enterprises.

[0006] In order to achieve the above object, the present invention provides a method for detecting electricity theft users in a low-voltage distribution network, comprising:

[0007] Obtain the user data to be detected and preprocess the user data to be detected.

[0008] The preprocessed user data to be detected is subjected to feature extraction to obtain a user feature data set to be detected.

[0009] The user feature data set to be detected is input into the pre-trained Wide and Deep model for detection to obtain the electricity theft users.

[0010] Furthermore, the preprocessing of the user data to be detected is specifically as follows:

[0011] Obtain the electricity consumption data of users in the area to be detected, and correct the data for extremely large outliers and null values.

[0012] Furthermore, feature extraction is performed on the preprocessed user data to be detected to obtain a feature data set of the user to be detected, specifically:

[0013] The preprocessed user data to be detected is input into the stack denoising autoencoder for data processing to obtain the user feature data set to be detected; wherein the stack denoising autoencoder includes a plurality of denoising autoencoders stacked together, the output of the previous denoising autoencoder is the input of the next denoising autoencoder, and the output of the last denoising autoencoder is used as the user feature data set to be detected.

[0014] Furthermore, the data processing step of the denoising autoencoder includes:

[0015] The pre-processed user data to be detected is recorded as data set G. Data set G is transformed into coding feature y through the encoding function h of the encoder. The conversion relationship between G and y is: y = θ (G) = s(WG + b); where s is the activation function of the neural network, θ = {W, b} is the parameter set, W is the parameter weight, and b is the bias value;

[0016] The reconstructed data of the data set G is calculated by the decoding function g: in, To reconstruct the data set, W′ is the weight coefficient, b′ is the bias coefficient, and θ′={W′,b′} is the parameter set of the decoding function g.

[0017] Furthermore, the optimization objective function of the stack denoising autoencoder is:

[0018] The parameters θ and θ′ are optimized with the goal of minimizing the value of the objective function. When the value of the objective optimization function is minimized, the y obtained is a user feature data set to be detected that obtains most of the features of the data set G.

[0019] Furthermore, the optimization objective function of the denoising autoencoder is

[0020] Among them, α is the reconstruction cost weight of the denoised dimension, β is the reconstruction cost weight of the undenoised dimension, G is the original input, G j is the jth element of the original data set G, n is the number of elements in the original data set G, z is the reconstructed data set, which is the reconstructed data obtained by randomly selecting a number of elements from the original data set G and setting them to 0, and calculating the encoding function and decoding function according to claim 4, z j is the jth element of the reconstructed dataset.

[0021] Furthermore, the training method of the Wide and Deep model includes:

[0022] Obtain daily electricity consumption data of low-voltage users with electricity theft tags and daily electricity consumption data of normal users in the model training area;

[0023] Correct the extremely large abnormal values ​​and null values ​​in the daily power consumption data of low-voltage users and normal users, and do not process the zero-value data of daily power consumption;

[0024] After data correction, we get the electricity stealing user data set and the normal user data set;

[0025] Inputting the electricity stealing user data set and the normal user data set into the stacked denoising autoencoder for feature extraction to obtain a first training feature set and a second training feature set;

[0026] The first training feature set and the second training feature set are input into the Wide and Deep model for training to obtain a model that can be used for electricity theft user detection.

[0027] Furthermore, the Wide and Deep model is specifically:

[0028] The Wide model is a logistic regression model. The Wide model expression is: y = w T x+b;

[0029] Among them, y is the predicted value, x=[x1,x2,…,x d ] are d eigenvectors, w=[w1,w2,…,w d ] is the model parameter, b is the prediction bias;

[0030] The Deep model is a deep neural network, which includes an input layer, a hidden layer, and an output layer.

[0031] Furthermore, the neurons in the input layer obtain information and input it into the hidden layer; the neurons in the hidden layer are used to process information, and Sigmoid is used as the activation function; the output of the hidden layer is in the form of: f(W1X+b1), where f is the Sigmoid function, and the complete deep neural network is f(x)=f(b2+(S(b1+W1x)), b2 and W2 represent the deviation and weight coefficient from the input layer to the hidden layer, respectively; b1 and W1 represent the deviation and weight coefficient from the hidden layer to the output layer, respectively.

[0032] The invention also discloses a device for detecting electricity theft users, comprising a preprocessing module, a feature extraction module and a detection module.

[0033] The preprocessing module is used to obtain the user data to be detected and preprocess the user data to be detected.

[0034] The feature extraction module is used to extract features from the preprocessed user data to be detected to obtain a user feature data set to be detected.

[0035] The detection module is used to input the user feature data set to be detected into a pre-trained Wide and Deep model for detection to obtain electricity theft users.

[0036] Compared with the prior art, the method and device for detecting electricity theft users in a low-voltage distribution network according to the embodiment of the present invention have the beneficial effect that, after obtaining the data of users to be detected in the area to be detected and inputting them into a preset Wide and Deep model after preprocessing and feature extraction, electricity theft users can be efficiently and accurately screened out, thereby ensuring the economic benefits of the power grid enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of a method for detecting electricity theft users in a low-voltage distribution network according to the present invention;

[0038] Figure 2 The present invention is a schematic structural diagram of a device for detecting electricity theft users. DETAILED DESCRIPTION

[0039] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0040] Therefore, many domestic and foreign scholars have begun to explore and study data-driven electricity theft detection methods, mainly through two technical routes: clustering methods that find data outliers and classifiers that learn existing labeled data for detection. Clustering methods are difficult to truly achieve accurate recognition in practical applications, and the main problem with classification methods is that there is no method that can simultaneously achieve a good fitting effect on the data cold start problem and have good generalization ability. At the same time, in feature engineering, it is necessary to find a method that can adapt to the high-dimensional characteristics of data to mine its deep features and solve the problem of too much extreme data affecting model fitting.

[0041] The decorrelation autoencoder cannot adapt well to the data characteristics of electricity theft users with a large number of extreme cases such as 0 values, and there is no single model in the classifier that can simultaneously take into account high fitting ability and strong generalization performance.

[0042] Embodiment 1:

[0043] like Figure 1 As shown, the present invention discloses a method for detecting electricity theft users in a low-voltage distribution network, which is applied to the detection of electricity theft users in the power grid, and the main steps include:

[0044] Step S1, obtaining user data to be detected and preprocessing the user data to be detected.

[0045] Step S2, extracting features from the preprocessed user data to be detected to obtain a user feature data set to be detected;

[0046] Step S3: input the user feature data set to be detected into the pre-trained Wide and Deep model for detection to obtain electricity theft users.

[0047] In step S1, the user data to be detected is obtained and pre-processed.

[0048] In this embodiment, the preprocessing of the user data to be detected is specifically as follows:

[0049] Obtain the electricity consumption data of users in the area to be detected, and correct the data for extremely large outliers and null values.

[0050] In this embodiment, since the obvious characteristic of many electricity stealing users is that a large number of daily electricity consumption values ​​are zero, the data with zero electricity consumption is not processed, and for extremely large abnormal values ​​and null values, the traditional quadratic interpolation method is used to correct the data.

[0051] In this embodiment, it should be emphasized that the amount of data from low-voltage users is large, and the minimum granularity data that can be obtained currently is their daily electricity consumption data. It is impossible to detect electricity theft through multi-dimensional data such as voltage and current like dedicated transformer users.

[0052] In the prior art, the data-driven electricity theft detection method mainly achieves electricity theft detection through two technical routes: clustering method for finding data outliers and classifier for learning existing labeled data for detection. Clustering method is difficult to truly achieve accurate recognition in practical applications, and the main problem with classification method is that there is no method that can simultaneously achieve a good fitting effect for data cold start problem and have good generalization ability. At the same time, in feature engineering, it is necessary to find a method that can adapt to the high-dimensional characteristics of data to mine its deep features and solve the problem of too much extreme data affecting model fitting.

[0053] In this embodiment, the decorrelation autoencoder cannot adapt well to the data characteristics of extreme cases such as the presence of a large number of zero values ​​in electricity theft users, and no single model in the classifier can simultaneously take into account high fitting ability and strong generalization performance.

[0054] In this embodiment, due to the high dimension of low-voltage user data in actual engineering applications, directly inputting the data into a classifier will cause the classifier to be unable to learn the deep features of the data.

[0055] Therefore, in the present invention, the data extraction method of the user to be detected is improved, see step S2 for details.

[0056] In step S2, feature extraction is performed on the preprocessed user data to be detected to obtain a user feature data set to be detected, specifically:

[0057] The preprocessed user data to be detected is input into the stack denoising autoencoder for data processing to obtain the user feature data set to be detected; wherein the stack denoising autoencoder includes a plurality of denoising autoencoders stacked together, the output of the previous denoising autoencoder is the input of the next denoising autoencoder, and the output of the last denoising autoencoder is used as the user feature data set to be detected.

[0058] In this implementation, the method of extracting features from the user data to be detected and the method of extracting features from the daily electricity consumption data of low-voltage users with electricity theft tags and the daily electricity consumption data of normal users during the Wide and Deep model training process are the same.

[0059] In this embodiment, the data processing steps of the denoising autoencoder include:

[0060] The preprocessed user data to be detected is recorded as data set G. Data set G is transformed into coding feature y through the encoding function h of the encoder. The conversion relationship between G and y is: y = h θ (G) = s(WG + b); where s is the activation function of the neural network, θ = {W, b} is the parameter set; W is the parameter weight, and b is the bias value;

[0061] The reconstructed data of the data set G is calculated by the decoding function g: s(W′y+b′); where, To reconstruct the data set, W′ is the weight coefficient, b′ is the bias coefficient, and θ′={W′,b′} is the parameter set of the decoding function g.

[0062] In this embodiment, the original data set is G, the encoding function is h, and the decoding function is g. G obtains the encoding feature y through the h function, and then obtains the reconstructed data set through the g function Therefore, θ′={W′,b′} is the parameter set of the decoding function g, W′ is the weight coefficient, and b′ is the bias coefficient.

[0063] In this embodiment, the encoding feature y is converted to obtain the reconstructed data x of G.

[0064] In this embodiment, the Sigmoid function is used as the excitation function.

[0065] In this embodiment, the optimization objective function of the stack denoising autoencoder is:

[0066] The parameters θ and θ′ are optimized with the goal of minimizing the value of the objective function. When the value of the objective optimization function is minimized, the y obtained is a user feature data set to be detected that obtains most of the features of the data set G.

[0067] In this embodiment, the optimization objective function of the denoising autoencoder is

[0068] Among them, α is the reconstruction cost weight of the denoised dimension, β is the reconstruction cost weight of the undenoised dimension, G is the original input, G j is the jth element of the original data set G, n is the number of elements in the original data set G, z is the reconstructed data set, which is the reconstructed data obtained by randomly selecting a number of elements from the original data set G and setting them to 0, and calculating the encoding function and decoding function according to claim 4, z j is the jth element of the reconstructed dataset.

[0069] In this implementation, the current single machine learning method cannot simultaneously meet the requirements of high fitting ability and high generalization ability required for this problem. Therefore, the Wide and Deep model structure of the present invention adopts a joint training mode of logistic regression model and DNN model. The training error of the entire model will be back-propagated to the two models at the same time to adjust the parameter weights and change the layout ratio. The two models complement each other by using their advantages. The wide model focuses on learning sparse features while the deep model is responsible for feature mapping to improve the generalization ability of the model.

[0070] In this embodiment, the training method of the Wide and Deep model includes:

[0071] Obtain the daily electricity consumption data of low-voltage users with electricity theft tags and normal users in the model training area.

[0072] The extremely large abnormal values ​​and null values ​​in the obtained daily electricity consumption data of low-voltage users and normal users are corrected, and the zero-value data of daily electricity consumption is not processed.

[0073] After data correction, we get the electricity stealing user data set and the normal user data set.

[0074] The electricity stealing user data set and the normal user data set are input into the stack denoising autoencoder for feature extraction to obtain the first training feature set and the second training feature set.

[0075] The first training feature set and the second training feature set are input into the Wide and Deep model for training to obtain a model that can be used for electricity theft user detection.

[0076] In this embodiment, the Wide and Deep model is specifically:

[0077] The Wide model is a logistic regression model. The Wide model expression is: y = w T x+b;

[0078] Among them, y is the predicted value, x=[x1,x2,…,x d ] are d eigenvectors, w=[w1,w2,…,w d ] is the model parameter and b is the prediction bias.

[0079] In this embodiment, T is the matrix transpose.

[0080] In this embodiment, the feature set includes the original input feature and the cross-product feature. The expression of the cross-product feature is: ki ∈{0,1}; where is a 0 / 1 Boolean variable, if the i-th feature is the k-th transformation Φ k If it is part of , it is 1, otherwise it is 0. It adds nonlinearity to the generalized linear model while capturing the interaction between binary features.

[0081] The Deep model is a deep neural network, which includes an input layer, a hidden layer, and an output layer.

[0082] In this implementation, the number of hidden layers can be selected according to actual needs. In order to better adapt to the high-dimensional characteristics of data, the present invention sets the number of hidden layers to 50.

[0083] In this embodiment, the neurons in the input layer obtain information and input the information into the hidden layer; the neurons in the hidden layer are used to process information, and Sigmoid is used as the activation function; the output form of the hidden layer is: f(W1X+b1), where f is the Sigmoid function, and the complete deep neural network is f(x)=f(b2+(S(b1+W1x)), b2 and W2 respectively represent the deviation and weight coefficient from the input layer to the hidden layer; b1 and W1 respectively represent the deviation and weight coefficient from the hidden layer to the output layer.

[0084] In this embodiment, the complete deep neural network includes three layers: input layer, hidden layer and output layer. The input sample of the input layer is X, and the mapping relationship from the input layer to the hidden layer is f(W1X+b1), that is, the input sample outputs f(W1X+b1) after passing through the hidden layer. Since f is a Sigmoid function, the output of the hidden layer is also expressed as S(W1X+b1). The hidden layer to the output layer is mathematically a softmax logistic regression process, and the function is expressed as f(W2X'+b2), where b2 and W2 represent the deviation and weight coefficient from the input layer to the hidden layer, respectively; X' represents the input of the output layer, that is, the output of the hidden layer. According to the above analysis, the functional relationship of the three layers of the input layer, hidden layer and output layer is combined, and the complete deep neural network can be obtained as f(x)=f(b2+(S(b1+W1x)). b1 and W1 represent the deviation and weight coefficient from the hidden layer to the output layer, respectively.

[0085] In this embodiment, Batch Normalization is introduced into the Deep model to solve the gradient vanishing problem. The input value of each hidden layer neuron is converted from a non-standard normal distribution to a normal distribution with a mean of 0 and a variance of 1 so that the gradient increases. The calculation formula is: Where x is the input vector, is x (k) The result after BN transformation.

[0086] In this embodiment, Dropout is introduced in the Deep model to solve the overfitting problem. By setting half of the hidden layer node values ​​to 0 in each training process, the network structure is simplified to weaken the co-adaptation relationship between the hidden layer nodes, thereby preventing the model from relying too much on certain local features during forward propagation, thereby obtaining stronger generalization ability.

[0087] In step S3, the user feature data set to be detected is input into a pre-trained Wide and Deep model for detection to obtain electricity theft users.

[0088] In this implementation, the detection model obtained after pre-training based on electricity theft users and normal users can quickly and accurately analyze and screen the data of users to be detected in the detection area to obtain electricity theft users. The power grid company can further conduct targeted screening based on the screened electricity theft users, which can greatly reduce the workload of screening electricity theft users and effectively protect the economic benefits of the power grid company.

[0089] Embodiment 2:

[0090] Reference Figure 2 The present invention discloses a device for detecting electricity theft users, comprising: a preprocessing module 101, a feature extraction module 102 and a detection module 103.

[0091] The preprocessing module 101 is used to obtain the user data to be detected and preprocess the user data to be detected;

[0092] The feature extraction module 102 is used to extract features from the pre-processed user data to be detected to obtain a user feature data set to be detected;

[0093] The detection module 103 is used to input the user feature data set to be detected into a pre-trained Wide and Deep model for detection to obtain electricity theft users.

[0094] In the present invention, Example 2 is written on the basis of Example 1, and includes the same technical features as Example 1, except that Example 2 protects the detection method of the present invention from the perspective of a device. The same technical features or limitations as those of Example 1 will not be repeated.

[0095] In summary, compared with the prior art, the method and device for detecting electricity theft users in a low-voltage distribution network according to the embodiment of the present invention have the following beneficial effects:

[0096] (1) After obtaining the data of users to be detected in the detection area, the data is preprocessed and feature extracted and input into the preset Wide and Deep model, which can efficiently and accurately screen out electricity theft users, thereby ensuring the economic interests of power grid companies.

[0097] (2) Using stacked denoising autoencoders to extract features from the dataset can not only effectively mine the deep features of the data itself, but also better deal with extreme situations such as a large amount of zero electricity consumption in the electricity theft user data.

[0098] (3) The Wide and Deep model is used as the classifier, which combines the strong fitting of logistic regression and is more effective in dealing with cold start problems; the DNN model has strong expressiveness and generalization ability and can effectively deal with electricity theft detection.

[0099] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention. These improvements and substitutions should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting electricity theft users in a low-voltage distribution network, characterized in that: include: Acquire the user data to be detected and pre-process the user data to be detected; Extracting features from the preprocessed user data to be detected to obtain a user feature data set to be detected; Input the user feature data set to be detected into the pre-trained Wide and Deep model for detection to obtain electricity theft users; The method further comprises: performing feature extraction on the preprocessed user data to be detected to obtain a feature data set of the user to be detected, specifically: inputting the preprocessed user data to be detected into a stacked denoising autoencoder for data processing to obtain a feature data set of the user to be detected; wherein the stacked denoising autoencoder comprises a plurality of stacked denoising autoencoders, the output of the previous denoising autoencoder is the input of the next denoising autoencoder, and the output of the last denoising autoencoder is used as the feature data set of the user to be detected; The optimization objective function of the stack denoising autoencoder is: The parameters θ and θ′ are optimized with the goal of minimizing the value of the objective function. When the value of the objective optimization function is minimized, the y obtained is the user feature data set to be detected that obtains most of the features of the data set G; The Wide and Deep model is specifically: The Wide model is a logistic regression model. The Wide model expression is: y = w T x+b; Among them, y is the predicted value, x=[x1,x2,…,x d ] are d eigenvectors, w=[w1,w2,…,w d ] is the model parameter, b is the prediction bias; The Deep model is a deep neural network, which includes an input layer, a hidden layer, and an output layer.

2. A method for detecting electricity theft users in a low-voltage distribution network according to claim 1, characterized in that: The preprocessing of the user data to be detected is specifically as follows: Obtain the electricity consumption data of users in the area to be detected, and correct the data for extremely large outliers and null values.

3. A method for detecting electricity theft users in a low-voltage distribution network according to claim 2, characterized in that: The data processing steps of the stack denoising autoencoder include: The preprocessed user data to be detected is recorded as data set G. Data set G is transformed into coding feature y through the encoding function h of the encoder. The conversion relationship between G and y is: y = h θ (G) = s(WG + b); where s is the activation function of the neural network, θ = {W, b} is the parameter set; W is the parameter weight, and b is the bias value; The reconstructed data of the data set G is calculated by the decoding function g: in, To reconstruct the data set, W′ is the weight coefficient, b′ is the bias coefficient, and θ′={W′,b′} is the parameter set of the decoding function g.

4. A method for detecting electricity theft users in a low-voltage distribution network according to claim 3, characterized in that: The optimization objective function of the denoising autoencoder is Among them, α is the reconstruction cost weight of the denoised dimension, β is the reconstruction cost weight of the undenoised dimension, G is the original input, G j is the jth element of the original data set G, n is the number of elements in the original data set G, and z is the reconstructed data set, which is obtained by randomly selecting several elements from the original data set G and setting them to 0, and calculating the reconstructed data through the encoding function and the decoding function. j is the jth element of the reconstructed dataset.

5. A method for detecting electricity theft users in a low-voltage distribution network according to claim 1, characterized in that: The training method of the Wide and Deep model includes: Obtain daily electricity consumption data of low-voltage users with electricity theft tags and daily electricity consumption data of normal users in the model training area; Correct the extremely large abnormal values ​​and null values ​​in the daily power consumption data of low-voltage users and normal users, and do not process the zero-value data of daily power consumption; After data correction, we get the electricity stealing user data set and the normal user data set; Inputting the electricity stealing user data set and the normal user data set into the stacked denoising autoencoder for feature extraction to obtain a first training feature set and a second training feature set; The first training feature set and the second training feature set are input into the Wide and Deep model for training to obtain a model that can be used for electricity theft user detection.

6. A method for detecting electricity theft users in a low-voltage distribution network according to claim 5, characterized in that: After obtaining information, the neurons in the input layer input the information into the hidden layer; the neurons in the hidden layer are used to process information, and Sigmoid is used as the activation function; the output of the hidden layer is in the form of: f(W1X+b1), where f is the Sigmoid function and X is the input sample of the input layer; the complete deep neural network is f(x)=f(b2+(S(b1+W1x))), where b2 represents the deviation from the input layer to the hidden layer; b1 and W1 represent the deviation and weight coefficient from the hidden layer to the output layer, respectively.

7. A device for detecting electricity theft users, characterized in that: include: Preprocessing module, feature extraction module and detection module; The preprocessing module is used to obtain the user data to be detected and preprocess the user data to be detected; The feature extraction module is used to extract features from the preprocessed user data to be detected to obtain a user feature data set to be detected; The detection module is used to input the user feature data set to be detected into the pre-trained Wide and Deep model for detection to obtain the electricity theft users; The method further comprises: performing feature extraction on the preprocessed user data to be detected to obtain a feature data set of the user to be detected, specifically: inputting the preprocessed user data to be detected into a stacked denoising autoencoder for data processing to obtain a feature data set of the user to be detected; wherein the stacked denoising autoencoder comprises a plurality of stacked denoising autoencoders, the output of the previous denoising autoencoder is the input of the next denoising autoencoder, and the output of the last denoising autoencoder is used as the feature data set of the user to be detected; The optimization objective function of the stack denoising autoencoder is: The parameters θ and θ′ are optimized with the goal of minimizing the value of the objective function. When the value of the objective optimization function is minimized, the y obtained is the user feature data set to be detected that obtains most of the features of the data set G; The Wide and Deep model is specifically: The Wide model is a logistic regression model. The Wide model expression is: y = w T x+b; Among them, y is the predicted value, x=[x1,x2,…,x d ] are d eigenvectors, w=[w1,w2,…,w d ] is the model parameter, b is the prediction bias; The Deep model is a deep neural network, which includes an input layer, a hidden layer, and an output layer.

Images