CN106952181A - Electric Load Prediction System based on long Memory Neural Networks in short-term - Google Patents

Abstract

The present invention discloses a kind of Electric Load Prediction System based on neural (LSTM) network of long short-term memory, and wherein LSTM networks include input layer, LSTM Internets and output layer, and the system includes：Information receiving module, for by the Power system load data of the historical juncture of input and provincial characteristics factor and being transferred to input layer；Model building module, for being trained modeling to the Power system load data and provincial characteristics factor of historical juncture by LSTM Internets, to generate deep neural network load forecasting model；Power prediction module, load forecast result in the region is produced for being predicted using deep neural network load forecasting model to the electric load in region, and by being connected to the recurrence device of LSTM Internets；As a result output module, for passing through the load forecast result in output layer output area.The present invention builds the load forecasting model of multi-task learning based on LSTM networks, can accurately predict the power load of multizone, improve prediction effect.

Description

Power load prediction system based on long-time and short-time memory neural network

Technical Field

The invention relates to the technical field of power load prediction, in particular to a power load prediction system based on a long-time and short-time memory neural network.

Background

The power load prediction problem aims at predicting the power consumption demand of a single or a plurality of power transmission lines in a power grid, and the power load prediction problem can be divided into the following steps according to the predicted time span: short term prediction (minutes to a week), medium term prediction (a month to a quarter), and long term prediction (more than a year). Under the condition of the prior art, electric energy is difficult to be effectively stored in a large-scale electric storage device, so that the residual generated energy is reduced as much as possible under the condition of meeting the power supply requirement, and the method is an effective way for reducing the cost and improving the use efficiency of the electric energy. Therefore, the method for accurately predicting the medium-short term power supply load in the area by adopting various prediction methods is necessary for planning and guiding the power generation enterprises to effectively produce electric energy. Currently, there are many mainstream methods applied to power load prediction, such as Artificial Neural Network (ANN), Support Vector Machine (SVM), Gaussian Process Regression (GPR), autoregressive Moving Average Model (ARIMA), and the like. The power load is associated with many hidden variables, such as lighting, wind, holidays, etc., which are generally difficult to obtain or quantify, but it is reasonable to consider cities in the same area to have similar hidden variables. The power load data of the adjacent cities are highly correlated, and the application of the multi-task learning technology can improve the load prediction accuracy of the similar areas.

The multi-task learning is a technology for improving generalization ability by simultaneously and jointly learning a plurality of related tasks, and when partial parameters in a model are reasonably shared among the tasks, the load prediction effect of the related tasks can be simultaneously improved. In recent years, with the deep development of deep learning theory research, it is a significant task to apply the deep learning theory to the prediction of the power demand of the power system. The existing various prediction methods based on the neural network can rarely predict the cross-regional power load at the same time, and the proposed power supply load prediction model is not accurate.

Disclosure of Invention

The invention mainly aims to provide a long-time and short-time memory neural network-based power load prediction system, which is used for constructing a multi-task learning load prediction model based on a long-time and short-time memory neural network (LSTM) in the deep learning field and can accurately predict the power loads of a plurality of adjacent areas at the same time.

In order to achieve the above object, the present invention provides an electrical load prediction system based on a long-term memory neural network, which is operated in a computer, the computer including an input unit and an output unit, the long-term memory neural (LSTM) network including an input layer, an LSTM network layer, and an output layer, the electrical load prediction system including:

the information receiving module is used for receiving the input power load data and the area characteristic factors at the historical moment through an input unit and transmitting the power load data and the area characteristic factors at the historical moment to an input layer of the LSTM network;

the model establishing module is used for importing the power load data and the regional characteristic factors at the historical moment received by the input layer of the LSTM network into the LSTM network layer, and training and modeling the power load data and the regional characteristic factors at the historical moment through the LSTM network layer so as to train and generate a deep neural network load prediction model, wherein the deep neural network load prediction model is a single-layer multi-task deep neural network model or a double-layer multi-task deep neural network model for power supply load prediction;

the power prediction module is used for predicting the power load in the region needing to be predicted by using the deep neural network load prediction model and generating a power load prediction result in the region through a regressor connected to the LSTM network layer;

and the result output module is used for outputting the power load prediction result in the required prediction area to the output unit through the output layer.

Preferably, the deep neural network load prediction model is expressed as the following formula:

Forecast＝f(t,d,c,y₁,u₁,id)

wherein, t ∈ [0,24]Is the time of day in hours, d ∈ {1, 2.., 365,366} is the number of days of the year in days, c is the type of day, y_lIs historical power load data containing a segment of historical power consumption requirements; u. of_lIs a real-valued vector containing regional characteristic factors; id represents the area identification of the electricity demand.

It is preferable thatThe LSTM network is an improved iterative neural network which is constructed by fitting hidden layer state vectors h_tRecursively applying a state transfer function f to process a network of input sequences, a hidden state vector h at a time step t_tFrom the current input sequence x_tAnd hidden state vector h of last moment_t-1Determining the hidden state vector h_tThe following formula is adopted:

preferably, the LSTM network layer includes an input gate i_tAnd an output gate o_tAnd forget door f_tAnd a memory cell c_tAt time t, memory cell c_tAll history information up to the current time t is recorded and input to the gate i_tAnd an output gate o_tAnd forget door f_tThe three logic gates are controlled, and the output values of the three logic gates are all between 0 and 1.

Preferably, the forgetting door f_tInformation erasing of controlling LSTM network layer, said input gate i_tControlling information update of LSTM network layer, said output gate o_tAnd controlling the information output of the internal state.

Preferably, the input sequence of the LSTM network is x ═ (x)₁,x₂,...,x_T) The input sequence is input into the LSTM network layer from the input layer, and the output sequence is y ═ y₁,y₂,...y_T) And outputting the parameters from the LSTM network layer by the output layer, wherein T is a prediction period, x is historical input data, and y is a predicted power load, and parameters of the LSTM network layer are updated iteratively according to the following formulas (1) to (6):

i_t＝σ(W_ix_t+U_ih_t-1+V_ic_t-1) (1)

f_t＝σ(W_fx_t+U_fh_t-1+V_fc_t-1) (2)

o_t＝σ(W_ox_t+U_oh_t-1+V_oc_t) (3)

h_t＝o_t⊙tanh(c_t) (6)

wherein x is_tThe method is characterized in that the method is an input sequence at the time t, sigma is represented by a sigmoid function, ⊙ is represented by multiplication among elements, W is an input weight, U is a cyclic weight of a hidden layer state h, V is an influence weight of historical information, and tanh is a hyperbolic tangent function of the hidden layer state h.

Preferably, a plurality of related tasks of the single-layer multi-task deep neural network model share one same LSTM network layer, and the output of the same LSTM network layer at the time t is represented asWherein the initialization parameters are uniformly distributed in [ -0.1, 0.1 [)]Random sample values in between.

Preferably, two related tasks of the double-layer multitask deep neural network model are respectively assigned to an LSTM network layer, each task respectively uses related information of the LSTM network layer of the other task, and information reception of the double-layer multitask deep neural network model is controlled through a global gating unit.

Preferably, the output of the LSTM network layer of the double-layer multitask deep neural network model at the time t is represented asAndwherein,andis uniformly distributed in [ -0.1, 0.1 ]]The (m, n) is a given group of related tasks, and the memory information of the LSTM network layer of the mth task is shown as the formula:

wherein,x_tthe method is characterized in that the method is an input sequence at the time t, sigma is represented as a sigmoid function, W is an input weight, U is a cyclic weight of a hidden state h, V is an influence weight of historical information, and tanh is a hyperbolic tangent function of the hidden state h.

Compared with the prior art, the Long-Short-term Memory Neural Network-based power load prediction system provided by the invention constructs a multi-task learning load prediction model based on the Long-Short-term Memory Neural Network (LSTM) in the deep learning field, so as to further improve the prediction effect. The invention provides a cross-regional power supply load prediction model, which can predict the power load of multiple regions at the same time, and has more accurate prediction effect compared with the existing power load prediction model.

Drawings

FIG. 1 is a block diagram of a preferred embodiment of a long-term memory neural network-based power load prediction system of the present invention;

FIG. 2 is a schematic diagram of a model architecture of an LSTM network;

FIG. 3 is a schematic diagram of a single-layer multitasking deep neural network model for power supply load prediction;

FIG. 4 is a schematic diagram of a two-layer multitasking deep neural network model for power supply load prediction.

The objectives, features, and advantages of the present invention will be realized and attained by the structure particularly pointed out in the written description and drawings.

Detailed Description

To further illustrate the technical means and effects of the present invention adopted to achieve the above objects, the following detailed description of the embodiments, structures, features and effects of the present invention will be made with reference to the accompanying drawings and preferred embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

Referring to fig. 1, fig. 1 is a block diagram of a preferred embodiment of the power load prediction system based on a long-and-short memory neural network according to the present invention. In the present embodiment, the electrical load prediction system 10 based on the long-term and short-term memory neural network is installed and operated in a computer 1, and the computer 1 further includes, but is not limited to, an input unit 11, a storage unit 12, a processing unit 13, and an output unit 14. The input unit 11 is an input device of a computer, such as an input keyboard or a mouse. The memory unit 12 may be a read only memory unit ROM, an electrically erasable and writable memory unit EEPROM, a FLASH memory unit FLASH or a solid hard disk. The Processing Unit 13 may be a Central Processing Unit (CPU), a Microcontroller (MCU), a data Processing chip, or an information Processing Unit with a data Processing function. The output unit 14 is an output device of the computer 1, such as a display or a printer.

The Long-Short Memory Neural Network-based power load prediction system 10 can construct a multi-task learning load prediction model based on a Long-Short Memory Neural Network (LSTM Network 2) in the deep learning field, so as to further improve the effect of regional power load prediction. Referring to fig. 2, the LSTM network 2 includes an input layer 21, an LSTM network layer 22, and an output layer 23.

In the present embodiment, the electrical load prediction system 10 based on a long-term and short-term memory neural network includes, but is not limited to, an information receiving module 101, a model building module 102, an electrical load prediction module 103, and a result output module 104. The module referred to in the present invention refers to a series of computer program instruction segments that can be executed by the processing unit 13 of the computer 1 and that can perform a fixed function, and that are stored in the storage unit 12 of the computer 1.

The information receiving module 101 is configured to receive, through the input unit 11, the input power load data and the area characteristic factors at the historical time, and transmit the power load data and the area characteristic factors at the historical time to the input layer 21 of the LSTM network 2; specifically, a typical power load demand prediction problem is influenced by various regional characteristic factors, the regional characteristic factors include information such as time, holidays, weather and economic indicators in a region, and the historical power load data refers to power load data information at a historical time in the region to be predicted. In the present embodiment, the historical power load data and the regional characteristic factors are collected by the user from the desired prediction region, and the information receiving module 101 receives the historical power load data and the regional characteristic factors from the input unit 11 and transfers them to the input layer 21 of the LSTM network 2.

The model establishing module 102 is configured to import the power load data and the regional characteristic factors received by the input layer 21 of the LSTM network 2 at the historical time into the LSTM network layer 2, and train and model the power load data and the regional characteristic factors at the historical time through the LSTM network layer 2 to train and generate a deep neural network load prediction model. In this embodiment, the deep neural network load prediction model is a single-layer multitask deep neural network model or a double-layer multitask deep neural network model for power supply load prediction. The model building module 102 performs load prediction modeling by using historical power load data and regional characteristic factors to generate a deep neural network load prediction model, which can be expressed as the following formula:

the variables in the above formula are illustrated as t ∈ [0,24 ]]Time of day in hours, d ∈ {1, 2.., 365,366} is the number of days of the year in days, c is the type of day, e.g., monday to sunday, holidays, etc., y is the number of days of the year_lIs a real-valued vector containing historical power load data within a segment of historical power demand; u. of_lThe method is a real value vector containing regional characteristic factors, such as temperature, economic index and other data; id represents the area identification of the electricity demand.

After the characteristic vectors are sampled and collected, a model can be constructed, namely a state transfer function f in the formula is determined, and then the electric load in a region is predicted. The present invention adopts an improved network long-time memory neural (LSTM) network of an iterative neural network (RNN) to perform modeling, and the structure and principle of the network model will be described in detail below.

An iterative neural network (RNN) is a network formed by fitting hidden state vectors h_tThe state transfer function f is applied recursively to process a network of arbitrarily long input sequences. Hidden state vector h at time step t_tFrom the current input x_tAnd hidden state vector h of last moment_t-1The determination is shown in the following formula:

the above formula can be regarded as a dynamic system, and the state of the system changes with time according to a certain rule. h is_tIs the state of the system, in theory, an iterative neural network (RNN) can approximate an arbitrary dynamic system. Traditionally, the strategy for modeling time series is to map the input sequence into a vector of fixed length using an iterative neural network (RNN) and then input it into a regressor, which gives the prediction result. However, in the training process of a plurality of RNNs based on the state transition function, after a long sequence is input, the gradient vector of the RNNs can grow or decay exponentially, which is the problem that the RNNs face gradient disappearance or gradient explosion. In this case, it is difficult for a plurality of RNNs to learn the long-term correlation problem of the sequence.

In this embodiment, the long-term memory neural (LSTM) network is an improved iterative neural network (RNN) model, and by introducing a logic gate mechanism, the problem of gradient disappearance or explosion faced by a simple iterative neural network is effectively solved, so that a deep network model can learn long-term dependence of a time sequence. The key to the LSTM network is the introduction of a set of Memory Units (Memory Units) that allow the network to learn when to forget historical information and when to update the Memory Units with new information.

As shown in fig. 2, fig. 2 is a model structural diagram of an LSTM network. In the present embodiment, the LSTM network 2 is composed of an input layer 21, an LSTM network layer 22 and an output layer 23, and the structure is shown in fig. 2. The LSTM network layer 22 includes an input gate i_t(input gate) and output gate o_t(output gate) and forget gate f_t(forkgate) and memory cell c_t. At time t, memory cell c_tAll history information up to the present time is recorded and controlled by three logic gates, which are: input door i_t(input gate) and output gate o_t(output gate) and forget gate f_t(forkgate). They can simulate the input, read and reset operations between neural cells, and the output values of the three logic gates are all between 0 and 1.

Let LSTM network 2 have x ═ x (x) as the input sequence₁,x₂,...,x_T) Input from the input layer 21 to the LSTM network layer 22, and the output sequence is y ═ y (y)₁,y₂,...y_T) And output from the LSTM network layer 22 by the output layer 23. Where T is the forecast period, x is historical input data (e.g., historical load, weather conditions, economic indicators, etc.), and y is the forecast load. To achieve this goal, the parameters of the LSTM network layer 22 are iteratively updated as shown in equations (1) - (6) below:

i_t＝σ(W_ix_t+U_ih_t-1+V_ic_t-1) (1)

f_t＝σ(W_fx_t+U_fh_t-1+V_fc_t-1) (2)

o_t＝σ(W_ox_t+U_oh_t-1+V_oc_t) (3)

h_t＝o_t⊙tanh(c_t) (6)

wherein x is_tIs an input sequence at the time t, sigma is expressed as a sigmoid function, tanh is a hyperbolic tangent function of a hidden state h, ⊙ is expressed as multiplication among elements, W is an input weight, U is a cyclic weight of the hidden state h, V is an influence weight of historical information, and the weight parameters are obtained through model training_tControl the information erasure in the LSTM network layer 22; input door i_tControls information updates in the LSTM network layer 22; output gate o_tControlling internal aspects of the LSTM network layer 22And outputting the information of the state.

In the present embodiment, the input gate i_tAnd an output gate o_tForgetting door f_tAnd a memory cell c_tCan make LSTM network layer 22 self-adaptively select forgetting, memorizing and outputting memorized information, if detecting important information content, forgetting gate f_tWill be turned off, and thus will utilize the information for a number of time steps, which is equivalent to capturing a long term dependency information; on the other hand, when forgetting to open the door f_tOn turn on, LSTM network layer 22 will choose to reset the memory state.

Most of the existing load prediction methods based on the neural network are in a single-task learning mode, and the methods are limited by the small number of training samples and cannot fully learn the network structure and parameters. To address this problem, these models incorporate an unsupervised pre-training phase. This unsupervised pre-training approach is effective to improve the final performance, but it is not a desirable task to directly optimize the system. The deep neural network model is well suited for multi-task learning because features learned from one task can be applied to improve learning of the remaining related tasks. The invention provides two deep neural network load prediction models based on a multitask learning framework, which are a single-layer multitask deep neural network model for power supply load prediction and a double-layer multitask deep neural network model for power supply load prediction respectively, wherein the specific model structures are shown in fig. 3 and fig. 4.

Referring to fig. 3 and 4, fig. 3 is a schematic diagram of a single-layer multitask deep neural network model for power supply load prediction; FIG. 4 is a schematic diagram of a two-layer multitasking deep neural network model for power supply load prediction. In fig. 3, multiple related tasks share one and the same LSTM network layer 22, and the output of this same LSTM network layer 22 at time t is represented asIn fig. 4, two related tasks are each assigned to one LSTM network layer 22,in this way, each task may use information about the LSTM network layer 22 of another task. It is worth noting that in fig. 4, given a set of related tasks (m, n), each with its own LSTM network layer 22, the output of the pair of LSTM network layers 22 at time t is represented asAndto better control the flow of shared information from one task to another, the present invention uses a global gating cell 31 to give the model the ability to decide how much information should be received. Based on the above equation (4), the memory content of the LSTM network layer 22 for the mth task is redefined as shown in equation (7):

wherein,the remaining parameter settings are consistent with the standard LSTM network layer 22, i.e.: x is the number of_tThe method is characterized in that the method is an input sequence at the time t, sigma is represented as a sigmoid function, W is an input weight, U is a cyclic weight of a hidden state h, V is an influence weight of historical information, and tanh is a hyperbolic tangent function of the hidden state h.

The power prediction module 103 is configured to predict a power load in a region to be predicted by using the deep neural network load prediction model, and generate a power load prediction result in the region through the regressor 30. According to the method, the power load in the area needing to be predicted can be predicted through the single-layer multitask deep neural network model or the double-layer multitask deep neural network model, and a power load prediction result in the area is generated. The two models provided by the invention can simultaneously and jointly learn two related tasks, the LSTM network layer 22 at the last layer of the model is connected with a Regressor 30, such as a Support Vector Regressor (Support Vector Regressor), and the predicted power load value can be output through the Regressor 30.

In the embodiment, the output of the LSTM network layer in the single-layer multitask deep neural network model at the time t is represented asWherein the initialization parameters are uniformly distributed in [ -0.1, 0.1 [)]Random sample values in between. In LSTM network layer 22 of the two-layer multitask deep neural network modelAndthe initialization parameters are uniformly distributed in [ -0.1, 0.1 [)]Random sample values in between. And training by using a minimum error sum of squares as a loss function and using an error back propagation algorithm, and searching for the hyper-parameters of the model by using a cross validation method experiment. The error back propagation algorithm and the cross validation method are both prior art in the technical field, and the present invention is not described in detail.

The result output module 104 outputs the prediction result of the power load in the required prediction area to the output unit 14 through the output layer 23; specifically, the output unit 14 outputs the prediction result of the power load in the region generated by the regressor 30, that is, the power load value of a group of related tasks (m, n) is y, through the output layer 23^(m)And y⁽ⁿ⁾。

Compared with the prior art, the invention has the following technical advantages: the short-term fluctuation information, the seasonal information and the trend information contained in the long-time load sequence can be simultaneously and jointly learned and stored, and the method is suitable for the multi-task high-dimensional time sequence prediction problem. The long-time and short-time memory neural network-based power load prediction system disclosed by the invention is used for constructing a multi-task learning load prediction model based on the long-time and short-time memory neural network (LSTM) in the deep learning field so as to further improve the prediction effect. The invention provides a cross-regional power supply load prediction model, which can predict the power load of multiple regions at the same time, and has more accurate prediction effect compared with the existing power load prediction model.

The above description is only a preferred embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent functions made by the present specification and drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.

Claims (9)

1. An electrical load prediction system based on a long-and-short-term memory neural network, which runs in a computer, the computer comprises an input unit and an output unit, the long-and-short-term memory neural (LSTM) network comprises an input layer, an LSTM network layer and an output layer, the electrical load prediction system comprises:

the information receiving module is used for receiving the input power load data and the area characteristic factors at the historical moment through an input unit and transmitting the power load data and the area characteristic factors at the historical moment to an input layer of the LSTM network;

the model establishing module is used for importing the power load data and the regional characteristic factors at the historical moment received by the input layer of the LSTM network into the LSTM network layer, and training and modeling the power load data and the regional characteristic factors at the historical moment through the LSTM network layer so as to train and generate a deep neural network load prediction model, wherein the deep neural network load prediction model is a single-layer multi-task deep neural network model or a double-layer multi-task deep neural network model for power supply load prediction;

the power prediction module is used for predicting the power load in the region needing to be predicted by using the deep neural network load prediction model and generating a power load prediction result in the region through a regressor connected to the LSTM network layer;

and the result output module is used for outputting the power load prediction result in the required prediction area to the output unit through the output layer.

2. The long-and-short-term memory neural network-based power load prediction system of claim 1, wherein the deep neural network load prediction model represents the following formula:

Forecast＝f(t,d,c,y₁,u₁,id)

wherein, t ∈ [0,24]Is the time of day in hours, d ∈ {1, 2.., 365,366} is the number of days of the year in days, c is the type of day, y_lIs historical power load data containing a segment of historical power consumption requirements; u. of_lIs a real-valued vector containing regional characteristic factors; id represents the area identification of the electricity demand.

3. The system of claim 1, wherein the LSTM network is a modified iterative neural network that operates by fitting hidden state vectors h_tRecursively applying a state transfer function f to process a network of input sequences, at timeHidden state vector h of step length t_tFrom the current input sequence x_tAnd hidden state vector h of last moment_t-1Determining the hidden state vector h_tThe following formula is adopted:

$h_t=\left\{\begin{array}{cc}0,& t=0\\ f(h_{t-1},x_t),& t\≠0\end{array}\right..$

4. the long-short memory neural network-based power load prediction system of claim 1, wherein the LSTM network layer comprises an input gate i_tAnd an output gate o_tAnd forget door f_tAnd a memory cell c_tAt time t, memory cell c_tAll history information up to the current time t is recorded and input to the gate i_tAnd an output gate o_tAnd forget door f_tThe three logic gates are controlled, and the output values of the three logic gates are all between 0 and 1.

5. The long-short term memory neural network-based power load prediction system of claim 4, wherein the forgetting gate f_tInformation erasing of controlling LSTM network layer, said input gate i_tControlling information update of LSTM network layer, said output gate o_tAnd controlling the information output of the internal state.

6. The long-short term memory neural network-based power load prediction system according to claim 4, wherein the input sequence of the LSTM network is x ═ (x ═ x₁,x₂,...,x_T) The input sequence is input into the LSTM network layer from the input layer, and the output sequence is y ═ y₁,y₂,...y_T) And outputting the parameters from the LSTM network layer by the output layer, wherein T is a prediction period, x is historical input data, and y is a predicted power load, and parameters of the LSTM network layer are updated iteratively according to the following formulas (1) to (6):

i_t＝σ(W_ix_t+U_ih_t-1+V_ic_t-1) (1)

f_t＝σ(W_fx_t+U_fh_t-1+V_fc_t-1) (2)

o_t＝σ(W_ox_t+U_oh_t-1+V_oc_t) (3)

${\tilde{c}}_t=\tanh (W_c{x}_t+U_c{h}_{t-1})---\left(4\right)$

h_t＝o_t⊙tanh(c_t) (6)

wherein x is_tThe method is characterized in that the method is an input sequence at the time t, sigma is represented by a sigmoid function, ⊙ is represented by multiplication among elements, W is an input weight, U is a cyclic weight of a hidden layer state h, V is an influence weight of historical information, and tanh is a hyperbolic tangent function of the hidden layer state h.

7. The electrical load prediction system based on a long-and-short memory neural network as claimed in claim 1, wherein a plurality of related tasks of the single-layer multitask deep neural network model share one same LSTM network layer, and the output of the same LSTM network layer at the time t is represented asWherein the initialization parameters are uniformly distributed in [ -0.1, 0.1 [)]Random sample values in between.

8. The electrical load prediction system based on an episodic memory neural network as claimed in claim 1, wherein two related tasks of the double-layer multitask deep neural network model are respectively assigned to an LSTM network layer, each task respectively uses the related information of the LSTM network layer of the other task, and the information reception of the double-layer multitask deep neural network model is controlled through a global gating unit.

9. The long-short term memory-based nerve of claim 8The system for predicting the power load of the network is characterized in that the output of the LSTM network layer of the double-layer multitask deep neural network model at the time t is expressed asAndwherein,andis uniformly distributed in [ -0.1, 0.1 ]]The (m, n) is a given group of related tasks, and the memory information of the LSTM network layer of the mth task is shown as the formula:

${\tilde{c}}^{\left(m\right)}=\tanh (W_c^{\left(m\right)}{x}_t+\underset{i\∈\{m,n\}}{\Σ}{g}^{(i\→m)}{U}_c^{(i\→m)}{h}_{t-1}^{\left(i\right)}),$

wherein,x_tthe method is characterized in that the method is an input sequence at the time t, sigma is represented as a sigmoid function, W is an input weight, U is a cyclic weight of a hidden state h, V is an influence weight of historical information, and tanh is a hyperbolic tangent function of the hidden state h.