Elevator energy consumption prediction method and device based on EEFG-LSTM-Attention
Through the neural network method based on EEFG-LSTM-Attention, the problem of difficult prediction of energy consumption peaks in elevator energy consumption prediction is solved, the prediction accuracy and performance are improved, and it is suitable for medium- and short-term prediction of elevator energy consumption.
Patent Information
- Application Number
- CN202211551376.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2022-10-21
- Filing Date
- 2022-12-05
- Publication Date
- 2025-01-03
- Estimated Expiration
- 2042-12-05
AI Technical Summary
In the prior art, when predicting elevator energy consumption, especially during the period of surge in people's flow, it is difficult to accurately predict the energy consumption peak, resulting in low prediction accuracy and the nonlinear characteristics of the medium and short-term energy consumption operation curves that degrade model performance.
Using the neural network method based on EEFG-LSTM-Attention, the training, verification and test data sets are constructed by acquiring and preprocessing elevator energy consumption data, and adjusting parameters through trial and compiling methods to build an energy consumption prediction model, using LSTM units and Attention The mechanism extracts energy consumption characteristics and laws to improve prediction accuracy.
It improves the identification accuracy and prediction performance of elevator energy consumption prediction, ensures operating speed, is better than the traditional LSTM, GRU and Bi-LSTM models, and can more accurately predict the energy consumption of elevators under future operating conditions.
Smart Images
Figure CN115935818B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of elevator energy consumption prediction, and in particular to an elevator energy consumption prediction method and device based on EEFG-LSTM-Attention. Background Art
[0002] Domestic energy consumption forecasting technology is mostly used in public buildings, power systems and other fields. In the field of forecasting technology, statistical methods or artificial intelligence methods are mostly used for scientific forecasting. In the statistical method, its analysis method is not sensitive to data with large changes. In the actual use of artificial intelligence, it depends more on the establishment of models and the selection of parameters. In general, the technology in the field of energy consumption forecasting in China is comprehensive and in-depth, whether it is traditional energy consumption forecasting based on physical models or artificial intelligence forecasting methods based on data-driven models. In the field of energy consumption forecasting abroad, machine learning technology is mostly used, and its technical focus is constantly shifting towards artificial intelligence, from early efforts to predict energy consumption analysis based on the characteristics of actual physical models, to training machine models with large amounts of data as the driving force. In the field of energy consumption forecasting, it is mostly used in building energy consumption.
[0003] In summary, for the prediction of elevator energy consumption in China, a hybrid prediction method based on ARMA and RBF neural network is adopted. The input data of this model must have the characteristics of stable time series. For the short-term elevator energy consumption operation curve, this research plan is feasible. However, the medium-term or longer time span elevator energy consumption operation curve has strong nonlinearity, and the prediction model performance will be reduced.
[0004] In view of the phenomenon that the surge in passenger flow in a short period of time leads to an increase in the fluctuation range of elevator energy consumption, the peak energy consumption is difficult to predict, resulting in low overall prediction accuracy. In energy consumption prediction, the fluctuation range of the energy consumption curve in a specific time period will be very large. Combined with actual application scenarios, when the elevator is in the upward peak operation mode, the energy consumption will surge, which is a difficulty for the prediction algorithm. The difference between the predicted value at the historical moment and the predicted value at the current moment is significantly different from their actual values. The difference between the peak prediction point and the actual value is huge, resulting in low peak data prediction performance. Summary of the invention
[0005] In order to solve the above technical problems, the present invention proposes an elevator energy consumption prediction method based on EEFG-LSTM-Attention.
[0006] In order to achieve the above object, the technical solution of the present invention is as follows:
[0007] The elevator energy consumption prediction method based on EEFG-LSTM-Attention includes the following steps:
[0008] Obtain elevator energy consumption data and preprocess it, and divide the preprocessed elevator energy consumption data into a training data set, a verification data set, and a test data set;
[0009] The training data set is input into the EEFG-LSTM-Attention neural network for training. The parameters are adjusted by trial and error to minimize the training error, and a preliminary elevator energy consumption prediction model is obtained. The preliminary elevator energy consumption prediction model is verified using the validation set to adjust the hyperparameters and start the next round of training until the number of iterations is reached.
[0010] The test data set is input into the preliminary elevator energy consumption prediction model for testing, and the prediction performance of the elevator energy consumption prediction model is evaluated and verified through the error index, and the prediction model with the best performance of each evaluation index is saved as the trained elevator energy consumption prediction model;
[0011] The elevator energy consumption data to be predicted is input into the trained elevator energy consumption prediction model, and energy consumption prediction is performed on the elevator energy consumption data to be predicted.
[0012] Preferably, the pretreatment comprises the following steps:
[0013] Normalize the elevator energy consumption data;
[0014] The normalized elevator energy consumption data is arranged in the form of an array and slid backward in sequence according to the size of the sliding window. The data contained in each sliding window is stored in a new array and put into the input data group. The data after each sliding window is put into the output data group, forming a pattern of input and output list pairs.
[0015] Preferably, the Min-Max standardization method is used to normalize the data, and its data conversion function is as follows:
[0016] X′=(XX min ) / (X max -X min )
[0017] Where X is the energy consumption value at a certain moment, X min , X max are the maximum and minimum values in the elevator energy consumption data set, and X' is the data value after data normalization.
[0018] Preferably, the elevator energy consumption prediction model includes an input layer, a hidden layer and an output layer, the number of neurons in the first three EEFG-LSTM unit layers in the hidden layer are 64, 128 and 256 respectively, which are used to extract the characteristics of energy consumption changes and related regularity information of the energy consumption data input by the input layer; the fourth layer, the Dropout layer, is used to randomly omit the connection between the output of the previous LSTM layer and the input of the next layer, and the parameter value is 0.1; the number of neurons in the fifth layer, the Attention layer, is 50, which is used to save the intermediate results generated by the time series input through the LSTM network, and to link the intermediate results with the output value to adjust the feature weight; the sixth and seventh layers are both fully connected layers, with the number of neurons being 50 and 1 respectively, which are used to output the predicted energy consumption value.
[0019] Preferably, the dimension of each sample data is adjusted to (samples, timestamp, features) as the input of the input layer, where samples represents the number of samples; timestamp is the time step, corresponding to the elevator energy consumption data at each time step in the sliding window; features is the input dimension, corresponding to the attribute of the elevator energy consumption data at a single time step, and the attribute is the average power, that is, features defaults to 1.
[0020] Preferably, the fully connected layer uses the ReLU function as the activation function.
[0021] Preferably, the ratio of the training data set to the test data set is 4:1, and 20% of the data in the training data set is used as the verification data set.
[0022] Preferably, during the model training process, the batch sample size is 64, the optimizer Adam is selected for model training, the learning rate is 0.01, and the number of iterations is 100.
[0023] Preferably, the error indicators include mean absolute error and mean square error.
[0024] Based on the above content, the present invention also discloses an elevator energy consumption prediction device based on EEFG-LSTM-Attention, including: an acquisition unit, a training unit, a testing unit and a prediction unit, wherein:
[0025] The acquisition unit is used to acquire and preprocess the elevator energy consumption data, divide the preprocessed elevator energy consumption data into a training data set and a test data set, and use 20% of the data in the training data set as a verification data set;
[0026] The training unit is used to input the training data set into the EEFG-LSTM-Attention neural network for training, adjust the parameters by trial and error to minimize the training error, and obtain a preliminary elevator energy consumption prediction model; use the verification set to verify the preliminary elevator energy consumption prediction model to adjust the hyperparameters, and start the next round of training until the number of iterations is reached;
[0027] The testing unit is used to input the test data set into the preliminary elevator energy consumption prediction model for testing, evaluate and verify the prediction performance of the elevator energy consumption prediction model through the error index, and save the prediction model with the best performance of various evaluation indicators as the trained elevator energy consumption prediction model;
[0028] The prediction unit is used to input the elevator energy consumption data to be predicted into the trained elevator energy consumption prediction model, and perform energy consumption prediction on the elevator energy consumption data to be predicted.
[0029] Based on the above technical scheme, the beneficial effects of the present invention are as follows: the present invention uses actual data as the driving force of the algorithm to predict the energy consumption of the elevator under actual future operating conditions, and provides an elevator energy consumption prediction model based on the use of differential forgetting long short-term memory unit (EEFG-LSTM) and attention mechanism (Attention-mechanism). Parameters are selected through trial and error, and the trained elevator energy consumption prediction model is compared with the LSTM, GRU and Bi-LSTM prediction models. Experiments are performed to improve the recognition accuracy while ensuring the operating speed. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 is a flow chart of an elevator energy consumption prediction method based on EEFG-LSTM-Attention in an embodiment;
[0031] Figure 2 is a schematic diagram of the sliding window concept in an embodiment;
[0032] Figure 3 is a schematic diagram of the structure of an elevator energy consumption prediction model in an embodiment;
[0033] Figure 4 is a schematic diagram of a three-dimensional array structure in an embodiment;
[0034] Figure 5 is an error tracking flow chart in one embodiment;
[0035] Figure 6 is a schematic diagram of the structure of an EEFG-LSTM unit in one embodiment;
[0036] Figure 7The figure is a block diagram of an elevator energy consumption prediction device based on EEFG-LSTM-Attention in an embodiment. DETAILED DESCRIPTION
[0037] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.
[0038] like Figure 1 As shown, this embodiment provides an elevator energy consumption prediction method based on EEFG-LSTM-Attention, and constructs an elevator energy consumption prediction model based on EEFG-LSTM-Attention neural network, which specifically includes the following steps:
[0039] Step S1, obtaining elevator energy consumption data and preprocessing it, dividing the preprocessed elevator energy consumption data into a training data set, a verification data set and a test data set;
[0040] In this embodiment, in order to enhance the fitting degree of the prediction curve and speed up the solution of the network model, a data normalization operation is used to limit the preprocessed data set to a certain range to solve the adverse effects caused by singular sample data. Specifically, the Min-Max normalization method is used to achieve data normalization, and its data conversion function is as follows:
[0041] X′=(XX min ) / (X max -X min )
[0042] Where X is the energy consumption value at a certain moment, X min , X max is the maximum and minimum value in the energy consumption data set, and X' is the data value after data normalization. Although this article is a time prediction of a single variable, there is no difference between data of different dimensions, but experiments have shown that the time difference between network model training after data normalization and before data normalization is relatively large. This is done to increase the speed of model training, speed up the convergence of the model, and reduce the error generated, so as to prevent the gradient explosion phenomenon caused by deep learning model training.
[0043] The energy consumption curve of the elevator is a time series. However, using machine learning to predict the curve is an input-to-output method, also known as supervised learning. The data set used in this article is a collection of two lists, one list corresponding to the input data, and the other list corresponding to the output data. The network model is learned from the two data columns. Therefore, before machine learning, the time series pattern needs to be converted into an input and output list pair pattern. This article uses a sliding window method to construct a supervised learning data set. The basic principle is: arrange the data set in the form of an array, slide backwards in sequence according to the window size, and each time the data contained in the window is re-stored in a new array and put into the input data group, and the data after the window is put into the output data group, thus forming an input and output list pair pattern. The size of the window number is called the lookback parameter, such as Figure 2 shown.
[0044] Finally, the data cutting step is performed. The machine learning data set is mainly divided into three parts: training data set, validation data set and test data set. The data set is divided into 80%, 20% for model training and testing, and 20% of the training data set is used for model verification.
[0045] Step S2: Input the training data set into the EEFG-LSTM-Attention neural network for training, adjust the parameters by trial and error to minimize the training error, and obtain a preliminary elevator energy consumption prediction model; use the validation set to verify the preliminary elevator energy consumption prediction model to adjust the hyperparameters, and start the next round of training until the number of iterations is reached.
[0046] Step S3, input the test data set into the preliminary elevator energy consumption prediction model for testing, evaluate and verify the prediction performance of the elevator energy consumption prediction model through error indicators, and save the prediction model with the best performance of various evaluation indicators as the trained elevator energy consumption prediction model.
[0047] In this embodiment, the elevator energy consumption prediction model based on the EEFG-LSTM-Attention neural network is constructed, which consists of: a data input layer, three EEFG-LSTM unit layers, a Dropout layer, an Attention layer, two fully connected layers and an output layer. The specific structure diagram of the prediction model is as follows Figure 3 shown.
[0048] The prediction model is explained as follows:
[0049] (1) Input layer
[0050] The input layer is a three-dimensional array, which is composed of (samples, timestamp, features), which represent the number of samples, time steps, and input dimensions respectively. The lookback parameter can be understood as the time step, with (x1,x2,x3,...,x n ) is the input data and the output value is predicted. The input dimension is represented by the energy consumption at a single time step with several features. In this paper, the data only has average power as a sample feature, so features defaults to 1, such as Figure 4 shown.
[0051] (2) Hidden Layer
[0052] The first three EEFG-LSTM unit layers in the hidden layer automatically learn the input energy consumption data, extract the characteristics of energy consumption changes and related regular information, and fit the prediction curve through a huge amount of parameters. In this paper, the first layer of EEFG-LSTM unit layer consists of 64 units; the second layer consists of 128 EEFG-LSTMD units; and the third layer consists of 256 EEFG-LSTM units. After three layers of learning and passing information to the next layer, each EEFG-LSTM unit includes an input gate, a forget gate, and an output gate. The error between the predicted value at the historical moment and the actual value at that moment is introduced into the forget gate of the next hidden layer unit. The information size of the forget gate is updated by the error size, thereby further determining the forgetting of historical information and correspondingly increasing the information size of the current input gate. The peak phenomenon of the surge in elevator energy consumption data can effectively improve the prediction accuracy, that is, dynamically adjust the historical information by transferring the error to improve the accuracy of the current prediction value.
[0053] Specifically, the long short-term memory unit LSTM (EEFG-LSTM) with differential forgetting is introduced. In the LSTM unit, reducing the effect of historical moment information on the current prediction is equivalent to reducing the LSTM unit state c(t-1) at time t-1, and this step can be achieved by changing the input of the forget gate. Therefore, the forget gate should adjust the input according to the deviation between the predicted value and the actual value at time t-1 to perform differential forgetting. For this purpose, the EEFG-LSTM uses the absolute error between the predicted value and the actual value at time t-1 and the memory c(t-1) of the previous moment as the input for updating the forget gate. This change takes into account the adjustment effect of the deviation between the predicted value and the actual value at time t-1 on the forget gate at time t, such as Figure 5 shown.
[0054] In the EEFG-LSTM unit structure, such as Figure 6As shown in the figure, the activation function of the forget gate is adjusted, and the sigmoid function in the standard LSTM is replaced with the ReLU function. The reason for this is that the derivative values of the sigmoid and tanh activation functions are less than 1, which easily leads to the gradient disappearance phenomenon.
[0055] In the formula, K = |h' t-1 -h t-1 |,h' t-1 is the actual value at time t-1; h t-1 is the predicted value at time t-1; C t-1 It is the memory state of the previous moment; ReLU is a new activation function that replaces sigmoid.
[0056] f t =ReLU(W f ·[K,C t-1 ]+b f )
[0057] i t =ReLU(W i ·[h t-1 ,x t ]+b i )
[0058] C t ′=tanh(W c [h t-1 ,x t ]+b c )
[0059] C t =f t ·C t-1 +i t ·C t '
[0060] O t =ReLU(W o ·[h t-1 ,x t ]+b o )
[0061] h t =O t tanh(C t )
[0062] The formula for the RELU activation function is:
[0063] ReLU(x)=max(0,x)
[0064] In the above formula, f t Forget gate, based on the K state at the previous moment and the input x at the current moment t, calculated with the memory state at time C, and adding the weight b c , and finally the result value is calculated by ReLU function.
[0065] i t Input gate, according to the state h at the previous moment t-1 and the current input x t , the same as the current memory state W i Calculate and add the weight b i , and finally the result value is calculated by ReLU function.
[0066] C` t The current memory output is based on the state h at the previous moment t-1 and the current input x t , the same as the current memory state W c Calculate and add the weight b c , and finally the result value is calculated by tanh function.
[0067] C t Memory output, multiply the memory output of the previous moment by the forget gate, and add the product of the input gate and the current memory output.
[0068] O t The calculation of the output gate is based on the state h at the previous moment t-1 and the current input x t , the same as the current memory state W o Calculate and add the weight b o , and finally the result value is calculated by ReLU function.
[0069] h t The value of the output gate O t Same as C calculated by tanh function t The product of .
[0070] h t =O t tanh(C t )
[0071] The fourth hidden layer is the Dropout layer, which is mainly used to randomly omit the connection between the output of the previous LSTM layer and the input of the next layer to prevent overfitting during training. In this paper, the parameter value of this layer is 0.1. Dropout is a commonly used method to prevent overfitting in machine learning.
[0072] The fifth hidden layer is the Attention layer. Multi-step inputs have different effects on the output, which affects the prediction results. The Attention mechanism is used to assign weights to inputs at different time steps to improve the energy consumption prediction effect. The main principle is to save the intermediate results generated by the time series input through the LSTM network, and link the results with the output values, so that the model learns how to selectively pay attention to the data and assign more reasonable weights to the data. The number of units is 50.
[0073] The sixth and seventh hidden layers are two fully connected layers (Dense Layer), with 50 and 1 units respectively. The number of units in the sixth layer depends on the training effect and the convergence speed of the model. Its main function is to extract features from the information transmitted by the previous layers. The correlation between energy consumption data can be extracted through nonlinear changes in the Dense layer, and finally mapped to the seventh layer. The seventh layer has only one unit because the problem solved by the prediction model is a regression problem, with the output value as the result, not the classification result, so a Dense unit is used to output the predicted energy consumption value.
[0074] (3) Output layer
[0075] The result of the output layer comes from the output of the seventh hidden layer, which is calculated and output by the ReLU activation function.
[0076] In the model construction, the number of hidden layer units and hidden layers has been determined. The batch sample size and training iteration number of the model training are 64 and 100 respectively. The batch sample size refers to the number of samples sent to the model each time the neural network is trained. In the LSTM neural network, a large batch size can make the network converge faster, but the convergence speed is also limited by hardware resources. Too large a batch may cause problems such as memory overflow. Generally, the value is usually 16, 32, 64, and 128. The number of iterations refers to the number of times the training set separated in the data set is input into the neural network for training. Its selection is mainly related to the curve fit, training standard, and test standard. In this paper, the loss, MAE value, RMSE value, and the intuitive actual-prediction curve effect are used to measure its value. Generally, the value is usually 50, 75, 100, and 125. After a large number of experiments, this paper takes the batch sample size of 64 and the iteration number of 100, and completes the parameter selection of the prediction network through trial and error. The parameters are shown in the following table.
[0077] Summary table of parameters of elevator energy consumption prediction model based on EEFG-LSTM-Attention neural network
[0078]
[0079] Step S4, inputting the elevator energy consumption data to be predicted into the trained elevator energy consumption prediction model, and performing energy consumption prediction on the elevator energy consumption data to be predicted.
[0080] Experimental verification
[0081] The performance differences of the four algorithm models in energy consumption prediction are reflected through experimental data. The maximum prediction error refers to the maximum value after subtracting the predicted data from the actual value under each group of sequences. The comparison is the reflection of the model in the worst prediction value. The average sample error and average sample deviation are derived from the MAE and RMSE evaluation criteria in the test set. The model prediction performance is measured from the perspective of the error value and deviation value of each group of sequences. The former measures the absolute error between data, and the latter measures the degree of dispersion between data. The minimum loss value is the minimum loss value during model training. The EEFG-LSTM-Attention model is superior to the other three in terms of the four performance parameters of maximum prediction error, average sample error, average sample deviation and minimum loss value. From the perspective of actual model understanding, the number of parameters based on the EEFG-LSTM-Attention model is not too complicated, and its running time is relatively the lowest. In the application process, the prediction speed can be improved, and the accuracy of data prediction is higher than that of the other three algorithm models. As shown in the following table.
[0082] Comparison of performance parameters based on EEFG-LSTM-Attention, LSTM, GRU and Bi-LSTM models
[0083]
[0084] It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.
[0085] In one embodiment, Figure 7 As shown, an elevator energy consumption prediction device 100 based on EEFG-LSTM-Attention is also provided, including: an acquisition unit 110, a training unit 120, a testing unit 130 and a prediction unit 140, wherein:
[0086] The acquisition unit 110 is used to acquire and preprocess the elevator energy consumption data, divide the preprocessed elevator energy consumption data into a training data set and a test data set, and use 20% of the data in the training data set as a verification data set;
[0087] The training unit 120 is used to input the training data set into the EEFG-LSTM-Attention neural network for training, adjust the parameters by trial and error to minimize the training error, and obtain a preliminary elevator energy consumption prediction model; use the verification set to verify the preliminary elevator energy consumption prediction model to adjust the hyperparameters, and start the next round of training until the number of iterations is reached;
[0088] The testing unit 130 is used to input the test data set into the preliminary elevator energy consumption prediction model for testing, evaluate and verify the prediction performance of the elevator energy consumption prediction model through the error index, and save the prediction model with the best performance of various evaluation indicators as the trained elevator energy consumption prediction model;
[0089] The prediction unit 140 is used to input the elevator energy consumption data to be predicted into the trained elevator energy consumption prediction model, and perform energy consumption prediction on the elevator energy consumption data to be predicted.
[0090] The devices and units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.
Claims
1. The elevator energy consumption prediction method based on EEFG-LSTM-Attention is characterized by: The steps include: Obtain elevator energy consumption data and preprocess it, and divide the preprocessed elevator energy consumption data into a training data set, a verification data set, and a test data set; The training data set is input into the EEFG-LSTM-Attention neural network for training, and the parameters are adjusted by trial and error to minimize the training error, so as to obtain a preliminary elevator energy consumption prediction model; the preliminary elevator energy consumption prediction model is verified by using the validation set to adjust the hyperparameters, and the next round of training is started until the number of iterations is reached. The elevator energy consumption prediction model includes an input layer, a hidden layer and an output layer. The number of neurons in the first three EEFG-LSTM unit layers in the hidden layer are 64, 128 and 256 respectively, which are used to extract the characteristics of energy consumption changes and related regularity information of the energy consumption data input in the input layer; The fourth layer, Dropout layer, is used to randomly omit the connection between the output of the previous LSTM layer and the input of the next layer. The parameter value is 0.
1. The number of neurons in the fifth layer, Attention layer, is 50. It is used to save the intermediate results generated by the time series input through the LSTM network, and to link the intermediate results with the output value to adjust the feature weight. The sixth and seventh layers are fully connected layers, with 50 and 1 neurons respectively, used to output the predicted energy consumption value; The test data set is input into the preliminary elevator energy consumption prediction model for testing, and the prediction performance of the elevator energy consumption prediction model is evaluated and verified through the error index, and the prediction model with the best performance of each evaluation index is saved as the trained elevator energy consumption prediction model; The elevator energy consumption data to be predicted is input into the trained elevator energy consumption prediction model, and energy consumption prediction is performed on the elevator energy consumption data to be predicted.
2. The elevator energy consumption prediction method based on EEFG-LSTM-Attention according to claim 1 is characterized in that: The pre-processing comprises the following steps: Normalize the elevator energy consumption data; The normalized elevator energy consumption data is arranged in the form of an array and slid backward in sequence according to the size of the sliding window. The data contained in each sliding window is stored in a new array and put into the input data group. The data after each sliding window is put into the output data group, forming a pattern of input and output list pairs.
3. The elevator energy consumption prediction method based on EEFG-LSTM-Attention according to claim 2 is characterized in that: The Min-Max standardization method is used to normalize the data, and its data conversion function is as follows: X′=(X-X min ) / (X max -X min ) Where X is the energy consumption value at a certain moment, X min , X max are the maximum and minimum values in the elevator energy consumption data set, and X' is the data value after data normalization.
4. The elevator energy consumption prediction method based on EEFG-LSTM-Attention according to claim 1 is characterized in that: The dimension of each sample data is adjusted to (samples, timestamp, features) as the input of the input layer, where samples represents the number of samples; timestamp is the time step, corresponding to the elevator energy consumption data at each time step in the sliding window; features is the input dimension, corresponding to the attribute of the elevator energy consumption data at a single time step, and the attribute is the average power, that is, features defaults to 1.
5. The elevator energy consumption prediction method based on EEFG-LSTM-Attention according to claim 1 is characterized in that: The fully connected layer uses the Relu function as the activation function.
6. The elevator energy consumption prediction method based on EEFG-LSTM-Attention according to claim 1 is characterized in that: The ratio of the training data set to the test data set is 4:1, and 20% of the data in the training data set is used as the verification data set.
7. The elevator energy consumption prediction method based on EEFG-LSTM-Attention according to claim 1 is characterized in that: During the model training process, the batch sample size is 64, the optimizer Adam is selected for model training, the learning rate is 0.01, and the number of iterations is 100.
8. The elevator energy consumption prediction method based on EEFG-LSTM-Attention according to claim 1 is characterized in that: The error indicators include mean absolute error and mean square error.
9. The elevator energy consumption prediction device based on EEFG-LSTM-Attention is characterized by: include: Get unit, training unit, test unit and prediction unit, where, The acquisition unit is used to acquire and preprocess the elevator energy consumption data, divide the preprocessed elevator energy consumption data into a training data set and a test data set, and use 20% of the data in the training data set as a verification data set; The training unit is used to input the training data set into the EEFG-LSTM-Attention neural network for training, adjust the parameters by trial and error to minimize the training error, and obtain a preliminary elevator energy consumption prediction model; use the verification set to verify the preliminary elevator energy consumption prediction model to adjust the hyperparameters, and start the next round of training until the number of iterations is reached. The elevator energy consumption prediction model includes an input layer, a hidden layer and an output layer. The number of neurons in the first three EEFG-LSTM unit layers in the hidden layer are 64, 128, and 256 respectively, which are used to extract the characteristics of energy consumption changes and related regularity information of the energy consumption data input by the input layer; the fourth layer Dropout layer is used to randomly omit the connection between the output of the previous LSTM layer and the input of the next layer, and the parameter value is 0.1; the number of neurons in the fifth layer Attention layer is 50, which is used to save the intermediate results generated by the time series input through the LSTM network, and link the intermediate results with the output value to adjust the feature weight; the sixth and seventh layers are both fully connected layers, with the number of neurons being 50 and 1 respectively, and are used to output the predicted energy consumption value; The testing unit is used to input the test data set into the preliminary elevator energy consumption prediction model for testing, evaluate and verify the prediction performance of the elevator energy consumption prediction model through the error index, and save the prediction model with the best performance of various evaluation indicators as the trained elevator energy consumption prediction model; The prediction unit is used to input the elevator energy consumption data to be predicted into the trained elevator energy consumption prediction model, and perform energy consumption prediction on the elevator energy consumption data to be predicted.