Energy time series data prediction system, platform and intelligent terminal device

Abstract

The application discloses an energy time series data prediction system, platform and intelligent terminal device. For the multivariate time series prediction of energy, the system collects real-time energy data through a data online collection module and constructs a time series, then extracts the latest multivariate time series segment through an input sequence extraction module when prediction is needed, and then predicts through a pre-trained multivariate time series prediction model. The energy time series data prediction system is universal for different forms of energy prediction. The application can provide a universal platform for power, natural gas and other energy data, and further provide prediction services. The platform can be combined with corresponding intelligent terminals to realize real-time release, query and visual display of prediction tasks.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of time series data prediction, and particularly relates to an energy time series data prediction system, platform and intelligent terminal device. BACKGROUND

[0002] In recent years, with the development of Internet of Things technology, various sensors have the ability to upload data regularly or in real time, generating a large amount of multi-element time series data composed of different sensors. As a monitoring device, the sequence data generated by the sensor is highly related to the energy equipment it monitors. For natural gas meters, photovoltaic power stations, electricity meters and other energy monitoring devices, they can generate time series data related to energy consumption or energy production through monitoring, and these energy-related data all have future

[0003] In the field of energy, there is often no natural physical connection between sensors, the use of the monitored device may change dynamically, and it has a high time correlation. The prediction model needs to be modeled according to the characteristics of the data to improve the prediction accuracy. The realization of energy data prediction can not only analyze the use or production rules of energy equipment, but also provide energy procurement reference for energy companies, reduce costs and generate economic benefits.

[0004] Energy time series data widely exists in different energy industries, such as natural gas, electricity, geothermal energy, etc. However, for different energy quantity prediction, there is no universal platform and model that is applicable to different prediction objects. Building a corresponding prediction platform and model for different industries on the one hand will cause repeated construction, and on the other hand will limit the development of some energy industries with weaker technical capabilities. SUMMARY

[0005] The purpose of the present application is to solve the problem that there is no universal platform and model that is applicable to different prediction objects in the prior art, and to provide an energy time series data prediction system, platform and intelligent terminal device.

[0006] The specific technical solutions adopted by the present application are as follows:

[0007] In a first aspect, the present application provides an energy time series data prediction system for predicting the energy quantity monitored by a plurality of energy quantity monitoring devices, the energy quantity being energy consumption or energy production, comprising:

[0008] a data online collection module for receiving real-time energy quantity data uploaded by each energy quantity monitoring device through the Internet of Things, and storing it in the form of time series with time stamp after associating it with the monitoring device identification number;

[0009] an input sequence extraction module configured to extract, according to a received prediction instruction, a latest energy amount data sequence of a fixed length time window from a time sequence stored by each energy amount monitoring device, and combine all the extracted sequence segments into a multivariate time sequence segment after normalizing each of the sequence segments;

[0010] an energy amount prediction module configured to input the multivariate time sequence segment generated by the input sequence extraction module into a pre-trained multivariate time sequence prediction model according to a received prediction instruction, and obtain a predicted value of the energy amount monitored by each energy amount monitoring device at a future specified time step.

[0011] As a preferred embodiment of the first aspect, the multivariate time sequence prediction model comprises a segmented graph convolution module, a time dilated convolution module, and a time information enhancement module, and the input of the model is a multivariate time sequence segment; the segmented graph convolution module segments the input multivariate time sequence segment according to the time dimension, constructs a segmented adjacency matrix according to the cosine similarity between the sequences in the segment, and then performs multi-layer segmented graph convolution to extract the correlation between the sequences, thereby obtaining a first output signal; the time dilated convolution module inputs the first output signal output by the segmented graph convolution module into a multi-layer stacked threshold control unit with dilated convolution, further extracts the features in each sequence and reduces the dimension, thereby obtaining a second output signal; the time information enhancement module inputs the input multivariate time sequence segment into an LSTM network, groups the signals of each hidden step of the LSTM network according to time, applies an attention mechanism to each group to extract a segmented context vector, then applies an attention mechanism on the segmented context vector to generate a time context vector, finally concatenates the time context vector with the last signal of the hidden layer of the LSTM network, and passes the concatenation result through a fully connected layer to obtain a third output signal; finally, the second output signal and the third output signal are added, and the result is de-normalized to obtain the final prediction output of the multivariate time sequence prediction model.

[0012] As a preferred embodiment of the first aspect, it further comprises an online learning module configured to continuously extract newly added data from the time sequence stored by each energy amount monitoring device, construct the data as a training sample of the multivariate time sequence prediction model, and retrain the multivariate time sequence prediction model using the training sample when the training sample accumulates to a specified number or a specified time has elapsed since the last model training, thereby realizing online learning of the multivariate time sequence prediction model.

[0013] As a preferred embodiment of the first aspect, in the input sequence extraction module, when normalizing the extracted sequence segment, the maximum value in the entire time sequence in which the sequence segment is located is used as a scaling value for normalization.

[0014] As a preferred form of the first aspect, two layers of graph convolution are used in the segment graph convolution module to extract the correlation between sequences; five layers of threshold control units with dilated convolution are used in the time dilation convolution module to extract features within each sequence.

[0015] As a preferred form of the first aspect, the energy quantity is the natural gas flow monitored by a natural gas flow metering device, or the electricity consumption monitored by an electricity meter device, or the photovoltaic power generation monitored by a photovoltaic power station.

[0016] In a second aspect, the present application provides an energy time series data prediction platform, which runs the energy time series data prediction system as described in any of the first aspect, and provides data access and instruction input interfaces to the outside.

[0017] As a preferred form of the second aspect, the prediction platform is a local server or a cloud platform.

[0018] In a third aspect, the present application provides a smart terminal device, which can establish a communication connection with the energy time series data prediction platform as described in the second aspect, and access data in the platform and send task instructions to the platform through the interfaces provided by the prediction platform.

[0019] As a preferred form of the third aspect, the task instructions sent by the smart terminal device to the platform include prediction instructions for executing prediction tasks, retraining instructions for calling online learning modules to retrain models, and specification instructions for specifying time steps required for prediction.

[0020] The present application has the following beneficial effects compared with the prior art:

[0021] 1) The present application provides an energy time series data prediction system for multi-element time series prediction of energy quantity. In the system, real-time energy quantity data is collected by a data online collection module and a time series is constructed, then the latest multi-element time series segment is extracted by an input sequence extraction module when prediction is needed, and then a pre-trained multi-element time series prediction model is used for prediction. The energy time series data prediction system is universal for different forms of energy quantity prediction.

[0022] 2) In the system of the present application, a multi-element time series prediction model based on deep learning is designed. The model constructs a segmented graph convolution layer, takes different time series as nodes in the graph, learns the correlation information of adjacent nodes from the graph structure based on sequence similarity, extracts the internal information of the time series through the time dilation convolution layer, and combines the output of the time information enhancement module based on the two-stage attention mechanism. The prediction experiment on the real energy quantity multi-element time series data set shows that the model can better complete the prediction of different energy quantities and different time steps, and accurately predict the energy quantity at different future times.

[0023] 3) The present application can provide a general platform for energy data such as electricity and natural gas, and further provide prediction services. The platform can be combined with the corresponding intelligent terminal to realize real-time release, query and visual display of the prediction task. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 FIG. 1 is a schematic diagram of the module composition of the energy time series data prediction system in an embodiment of the present application.

[0025] Figure 2 FIG. 2 is a schematic diagram of the information aggregation process of a single-layer graph neural network.

[0026] Figure 3 FIG. 3 is a schematic diagram of the network structure of the multi-element time series prediction model.

[0027] Figure 4 FIG. 4 is a schematic diagram of the module composition of the energy time series data prediction system in another embodiment of the present application.

[0028] Figure 5 FIG. 5 is a schematic diagram of the relationship between the energy quantity monitoring device, the energy time series data prediction platform and the intelligent terminal device in the embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings. In the following description, a large number of specific details are set forth in order to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the scope of the present application, so the present application is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present application can be combined accordingly without conflict.

[0030] In the description of the present application, it should be understood that when one element is considered to be "connected" to another element, it can be directly connected to the other element or indirectly connected to the other element with an intermediate element. In contrast, when an element is considered to be "directly" connected to another element, there is no intermediate element.

[0031] The present application provides an energy time series data prediction system for predicting the energy quantity monitored by a plurality of energy quantity monitoring devices. Figure 1 As shown in the preferred embodiment of the present application, the energy time series data prediction system comprises a data online collection module, an input sequence extraction module and an energy quantity prediction module.

[0032] The data online collection module is used to receive real-time energy quantity data uploaded by each energy quantity monitoring device through the Internet of Things, and store it in the form of a time series with a timestamp after associating it with the monitoring device identification number.

[0033] The input sequence extraction module is used to extract the latest energy quantity data sequence from the time series stored by each energy quantity monitoring device according to a fixed length time window according to the received prediction instruction, and combine all the extracted sequence segments into a multivariate time series segment after normalizing each of them.

[0034] The energy quantity prediction module is used to input the multivariate time series segment generated by the input sequence extraction module into a pre-trained multivariate time series prediction model according to the received prediction instruction, to obtain the energy quantity prediction value monitored by each energy quantity monitoring device at a future specified time step.

[0035] In the present application, the energy time series data prediction system described above is not limited to the energy quantity prediction object it faces. The energy quantity in the present application can refer to energy consumption or energy generation. In subsequent embodiments, the energy consumption can be natural gas consumption or electricity consumption, and the energy generation can be photovoltaic power generation. Therefore, different energy quantity monitoring devices are required for different energy quantities. For example, when the energy quantity is natural gas flow, it can be monitored by a natural gas flow metering device, when the energy quantity is electricity consumption, it can be monitored by an electric meter device, and when the energy quantity is photovoltaic power generation, it can be monitored by a photovoltaic power station. However, in addition to the above three energy quantities, the present application can also be applied to other energy-related time series data, such as wind power generation data, oil and gas production data, hydroelectric power generation data, thermal power generation data, etc.

[0036] The data transmission technology of each energy quantity monitoring device can be any Internet of Things technology that can remotely transmit data. In the embodiments of the present application, the following technologies can be used:

[0037] 1. Wi-Fi technology: Wi-Fi technology is a kind of local area network wireless transmission technology, which can provide high-speed and stable network connection. By connecting the device to the Wi-Fi network, remote control and data transmission of the Internet of Things device can be realized, which is suitable for small range and high-speed transmission scenarios.

[0038] 2. Cellular network technology: Cellular network technology is one of the core technologies in the field of mobile communication, which realizes remote data transmission by using 3G or 4G network, and has wide coverage and efficient and stable transmission characteristics. It can be applied to cross-regional and long-distance large-flow data transmission scenarios.

[0039] 3. Zigbee technology: Zigbee technology is a kind of wireless low-power Internet of Things technology, which can provide local area network level data transmission, and is suitable for low-speed and low-power transmission scenarios such as smart home, etc.

[0040] 4. LoRaWAN technology: LoRaWAN technology uses long-distance and low-power wireless communication technology, which can realize remote data transmission and monitoring in different coverage and environmental conditions in cities, villages and other places.

[0041] Of course, in addition to the above-mentioned several Internet of Things technologies, other Internet of Things technologies that can remotely transmit data can also be considered, as long as they can meet the timeliness and reliability of data transmission.

[0042] In the above data online collection module, the real-time energy quantity data uploaded by each energy quantity monitoring device through the Internet of Things can be recorded in the data storage module with the non-repeated device identification number as the unique identifier for subsequent calling. The data uploaded by each energy quantity monitoring device can be recorded in the same time sequence, and each record value needs to have a sampling timestamp.

[0043] In the present application, the function of the above-mentioned input sequence extraction module is to extract the input data required for the subsequent multivariate time series prediction model, i.e. multivariate time series segment. The extraction of the multivariate time series segment can be realized by sliding the time window. The multivariate time series prediction model of the present application is used to predict the future monitoring values of multiple energy quantity monitoring devices simultaneously, so the latest sequence in the time series corresponding to each energy quantity monitoring device needs to be extracted by the time window, and these sequences are arranged side by side to form a multivariate time series segment.

[0044] Therefore, the multivariate time series prediction model is the key to realizing the prediction function of the present application. The specific implementation of the multivariate time series prediction model is described in detail below. In the multivariate time series prediction task, an N-dimensional multivariate time series sample is generated by N variables changing over time, and each variable can correspond to an energy quantity monitoring device. In the graph neural network, the model regards each variable as a node in the graph and regards the correlation degree between the variables as an edge in the graph structure. The graph G k ={V,E} represents the structure of the graph, where V and E represent the set of nodes and edges, respectively. For the graph edge set E, u,v∈V are nodes in the graph, and e=(u,v)∈E represents an edge from node u to node v. The neighborhood of node u is defined as N(u)={v∈V|u,V)∈E}. The connectivity of the entire graph is represented by the adjacency matrix A∈R {N×N} , where A i,j =c>0 when there is a connected edge between node i and node j, otherwise A i,j =0. The main diagonal of the adjacency matrix A is 1, indicating that each node has a self-connection edge. The main function of the graph neural network is to enable each node to learn the mutual dependence of adjacent nodes in the graph while maintaining its own features, automatically aggregate information, and generate a new time series. The new time series output by the node contains hidden information of other time series. With the increase of the number of convolution layers, a single node will also indirectly obtain node information from a larger range of its adjacent nodes. As shown in Figure 2 , the information aggregation process of a single-layer graph neural network is shown: the center node (i.e., the energy quantity monitoring device 2) aggregates information from its neighbor nodes (i.e., the energy quantity monitoring devices 1, 3, and 4) and outputs a new time series.

[0045] Based on the above concept, as shown in Figure 3 , in an embodiment of the present application, the multivariate time series prediction model includes a segmented graph convolution module, a time dilation convolution module, and a time information enhancement module, and the input of the model is a multivariate time series segment. Assuming that the input of the model is the aforementioned multivariate time series segment X, the specific implementation of the three modules is as follows:

[0046] (1) The segmented graph convolution module segments the input multivariate time series segment according to the time dimension, constructs a segmented adjacency matrix according to the cosine similarity between the sequences in the segment, and then performs multi-layer segmented graph convolution to extract the correlation between the sequences and obtain a first output signal. Specifically, the multivariate time series segment X input into the model is input into the segmented graph convolution module to obtain the final graph convolution result H (G) . The process performed in the segmented graph convolution module is shown in S11-S14 as follows:

[0047] S11. First, based on the preset segmentation number hyperparameter P, the input multivariate time series segment X is divided into P segments in the time dimension on an average basis. Each segment is represented as X. k ∈R {×S} S = / , W is the length of the time window, and N is the total number of energy monitoring devices that need to be predicted.

[0048] S12, then segmented X k The sequences corresponding to each energy monitoring device within the system are nodes, and cosine similarity is used to calculate segmented X. k The adjacency degree of the internal nodes is used to obtain the segment X. k The corresponding graph adjacency matrix A′ k ∈R {N×N} A′ k Any node i and node j are in the same segment X k The strength of the connecting edge is expressed as:

[0049]

[0050] In the formula: X k,i and X k,j Representing segment X respectively k The sequence corresponding to the i-th and j-th energy monitoring devices, where || represents the modulo operation.

[0051] S13. Then use the Softmax function to modify the graph adjacency matrix A′. k Normalize by row to form the normalized graph adjacency matrix A. k A k In and A′ k,i,j The corresponding adjacency degree A k,i,j The calculation formula is:

[0052]

[0053] Since the range of cosine similarity is [-1, 1], a value closer to 1 indicates that the two vectors are more similar in direction, while a value less than 0 indicates that they are in opposite directions. Therefore, the above model sets 0 as the lower limit to filter irrelevant nodes when calculating the strength of connecting edges, and uses the Softmax function to apply the similarity matrix A of each multivariate time series segment. k Normalize line by line.

[0054] S14. Finally, for each segment X... k Based on the obtained segmented adjacency matrix A k The graph convolution network performs a total of G layers of graph convolution computation, and outputs the final graph convolution result H. (G) Furthermore, the graph convolution process of the l-th layer in the graph convolutional network is represented as follows:

[0055]

[0056] where D k is the degree matrix of the adjacency matrix A k , is a trainable parameter matrix, is the k-th segment of the final output of the previous graph convolution layer after decomposition; l = 1, 2, …, G,

[0057] The graph convolution network obtains the segment convolution result of the l-th layer of graph convolution reorganizes and splices in the time dimension to obtain the output of the l-th layer of graph convolution:

[0058]

[0059] The graph convolution network obtains the l-th layer of graph convolution result and the original input multivariate time series segment X to obtain the final output of the l-th layer of graph convolution The introduction of residual connection is to prevent gradient disappearance and make the model easier to train.

[0060] Through the multi-layer graph convolution process, each node in H G collects associated information from neighbor nodes and then passes it to the time dilation convolution layer.

[0061] (2) The time dilation convolution module inputs the first output signal output by the segment graph convolution module into a multi-layer stacked gate linear unit with dilation convolution, further extracts features within each sequence and reduces dimension to obtain a second output signal. Specifically, the final graph convolution result H (G) is input into the time dilation convolution module to perform a total of C layers of time dilation convolution to obtain the final output H C of the time dilation convolution module. Stack C layers of time dilation convolution layers, and each layer of time dilation convolution layer is a gate linear unit (GLU) with dilation convolution, which includes a feature learning part and a signal quantity control part, can learn the time variation law of each variable, and each convolution layer is divided into two signal paths: filter and signal threshold. The feature learning part learns the hidden features of the signal through linear weighting. The signal quantity control part uses a sigmoid activation function to control the signal quantity of the current convolution layer output entering the next layer. The specific process in the time dilation convolution module is shown in the following S21-S22:

[0062] S21, sequentially perform C layers of time dilation convolution, wherein the process of any l-th layer of time dilation convolution is represented as follows:

[0063]

[0064] where: c = 1, 2, …, C; Z (c-1) is the output of the previous dilated convolution layer, denotes matrix point multiplication (i.e. Hadamard product), σ(·) denotes sigmoid activation function, * denotes dilated convolution operation, W f (c) and denote the learnable parameters of the two-part convolution kernel respectively, the number of two-part convolution kernels is M, the size of the convolution kernel is C1, and the dilation rate is D. Since most time series have a relatively fixed sampling frequency (such as 1h or 10min) and periodicity at different time scales, the dilated convolution here can use 3*2 n as the dilation rate, so that the receptive field of the convolution kernel better covers the time scales from small to large (i.e. when the dilation rate is 24 (3*2 n ), the sampling frequency is 1h, and the receptive field of the convolution kernel will cover the same hour for consecutive days).

[0065] S22, get the signal Z (c) output by each layer of time dilated convolution (c) , and then use a layer of ordinary convolution to process the signal to the size of the prediction target (i.e. the prediction output size of the multivariate time series prediction model), so that the final output of the time dilated convolution module is:

[0066]

[0067] where W s (c) denotes the trainable weight matrix when weighting the signal Z (c) output by the cth layer of convolution; W e (·) represents ordinary convolution, the number of convolution kernels is M e , and the size of the convolution kernel is C e . H C Stacking the outputs of convolutions with different expansion rates can capture sequence information with different receptive fields.

[0068] (3) The time information enhancement module inputs the input multivariate time series segment into the LSTM network, groups the signals of each hidden step of the LSTM network according to time, and applies attention mechanism to each group to extract segment context vectors, then continues to apply attention mechanism on the segment context vectors to generate a time context vector, finally concatenates the time context vector with the last signal of the LSTM network hidden layer, and passes the concatenation result through a layer of fully connected layer to obtain the third output signal.

[0069] Specifically, the multivariate time series segment X of the input model is input into the time information enhancement module to obtain the time information enhancement signal H. A The execution flow in the time information enhancement module is shown in S31 to S34 below:

[0070] S31. First, feature extraction is performed using an LSTM network with N hidden layer units. Then, the hidden layer output H′ of the LSTM network with length W is segmented according to the segmentation number hyperparameter P, resulting in P segmented hidden layer signals H′. k :

[0071] H′=LSTM(X)=[h1,h2,…,h W ]

[0072] H′ k =[h k*s-s+1 ,h k*S-S+2 ,…,h k*S ], k=1,2,…P

[0073] S32. Then, the obtained segmented hidden layer signal H′ k After trainable parameters W k After weighting, we get The corresponding segment weights a′ are obtained by calculating using the Softmax function. k The hidden layer signal H′ of each segment k With the weights a′ of each segment k The weighted summation yields the piecewise context vector C′:

[0074]

[0075]

[0076]

[0077] S33. Then, the segmented context vector C′ is weighted by the trainable parameter W′′ to obtain... The hidden layer signal contains S time steps. The time step weights a″ are obtained after calculation using the Softmax function. i The time context vector C″ is composed of the hidden layer signals at each time step. With time step weight a″ i Weighted summation yields:

[0078]

[0079]

[0080]

[0081] S34, finally, the obtained time context vector C" is combined with the last time step signal h of the LSTM network hidden layer W concatenated, and output time information enhancement signal after a full connection layer:

[0082] H A =W A ·concat(C″,h W )+b

[0083] wherein: W A and b represent the learnable weights and bias of the full connection layer, respectively.

[0084] (4) Finally, the second output signal and the third output signal are added, and the final prediction output of the multivariate time series prediction model is obtained after inverse normalization.

[0085] Specifically, the final output H C of the time dilation convolution module and the time information enhancement signal H A output by the time information enhancement module are added, and the final prediction output of the network model is obtained after inverse normalization:

[0086]

[0087] wherein: scale represents the inverse normalization coefficient.

[0088] It should be noted that the inverse normalization here corresponds to the normalization performed on the input data, and therefore the inverse normalization coefficient scale corresponding to the prediction value of each energy quantity monitoring device is the maximum value of the corresponding time series recorded by the energy quantity monitoring device.

[0089] Some hyperparameters in the above multivariate time series prediction model can be optimized according to actual conditions. In the embodiment of the present application, the length W of the time window is 168, the sliding step s of the time window is 1; the segmentation number hyperparameter P is 7; the number of graph convolution layers G in the graph convolution network is 2; the number of time dilation convolution layers C in the time dilation convolution module is 5, the number of convolution kernels M of each layer is 24, the size of the convolution kernel C1 is 1x3, and the dilation rate D of the 5-layer convolution kernel is 1, 3, 6, 12, and 24, respectively; in the time dilation convolution module, the number of convolution kernels M e of the last ordinary convolution layer is 1, and the size of the convolution kernel C e is 1x76. It should be noted that the above hyperparameters are only preferred parameters of the embodiment of the present application, and are not absolute.

[0090] It should be noted that the above multi-element time series prediction model needs to be trained in advance before application, and the specific training method belongs to the prior art. The loss function used in the training can use the mean absolute error (MAE). During specific training, the final prediction output of the multi-element time series prediction model The mean absolute error is calculated with the predicted target actual label value Y of the input sample, so as to obtain the loss function value L:

[0091]

[0092] Since Y has been normalized in advance, it needs to be denormalized before MAE calculation.

[0093] Before training the above model, a training data set needs to be constructed in advance, which belongs to the prior art. For the convenience of understanding, a brief introduction is made below, and the process can be referred to as follows:

[0094] 1) Obtain the time series recorded by the N energy monitoring devices required to be monitored by the user Each time series The maximum value is taken as the scale value for maximum value normalization to form a normalized time series All normalized time series are combined to form a multi-element time series

[0095] The above maximum value normalization belongs to the prior art. Assuming that the maximum value in each time series is MAX(S[i]), it can be taken as a scale value to perform maximum value normalization on each sequence respectively:

[0096]

[0097] Among them represents the i-th time series, represents the i-th normalized historical time series.

[0098] 2) A time window with a fixed length W is slid on the multi-element time series according to a preset step s. During the sliding process, each time window extracts a multi-element time series segment X ∈ R {N×W} from the multi-element time series within the window, and takes the value Y ∈ R {N} at the h-th time step after the time window in the multi-element time series as the regression label of the model, to generate the training data set of the model. Each sample in the data set is represented as {X ∈ R {N×W} , ∈ R {N}}.

[0099] After obtaining the training data set, the model can be trained, and the random gradient of the model can be calculated according to the obtained prediction loss L. The AdamW optimizer is used to update all trainable parameters of the neural network model with a learning rate a and a decay weight b. The training process is repeated until the model converges, and the model is saved in the energy prediction module. Thus, in actual prediction, the sequence segment in the current nearest time window is obtained from the time series data recorded by each energy monitoring device and constitutes a real-time multivariate time series segment. After normalization, the real-time multivariate time series segment is input into the trained multivariate time series prediction model stored in the energy prediction module, and the recorded value corresponding to each energy monitoring device at the specified time step in the future is predicted.

[0100] In an embodiment of the present application, in order to show the prediction performance that the above-mentioned multivariate time series prediction model can achieve, it is tested on natural gas data sets (Nature-gas-1 and Nature-gas-2), solar energy data set (Solar-energy) and electricity data set (Electricity). The root relative squared error (Root Relative Squared Error, RSE) and the empirical correlation coefficient (Empirical Correlation Coefficient, CORR) are selected as the evaluation indicators of the model performance. The specific conditions of several data sets are as follows:

[0101] Natural gas data set (Nature-gas-1 and Nature-gas-2): Both data sets come from the commercial and industrial natural gas flow meter database of a gas company. The standard cumulative flow of the flow meter is resampled, interpolated and differentiated in advance to generate the natural gas standard flow sequence. Since different natural gas sensors have different upload frequencies, in order to show the performance of the multivariate time series prediction model under different sampling frequencies, two natural gas data sets with different sampling frequencies are constructed. Among them, Natural-gas-1 contains 52 high-frequency natural gas flow sequences from July 1, 2020 to September 5, 2021, with a sampling frequency of 10 min; Natural-gas-2 contains 118 low-frequency natural gas flow sequences from March 9, 2020 to September 5, 2021, with a sampling frequency of 1H. Due to the particularity of the natural gas industry, some commercial and industrial users are in a suspended gas state before and after the Spring Festival, so the data of February 2021 in the two data sets is deleted.

[0102] Solar energy data set (Solar-energy): contains solar energy production records from 137 photovoltaic power plants in Alabama in 2006, with a sampling frequency of 10 min.

[0103] Electricity: from UCI machine learning dataset, containing electricity usage of 321 customers from 2012 to 2014. The original data has a sampling frequency of 15min, which is converted into hourly electricity usage data here.

[0104] Meanwhile, in order to predict the performance of different prediction steps, the prediction steps of the multivariate time series prediction model in the embodiment are set to 4 groups, that is, h = 3, 6, 12, and 24, and experiments are carried out respectively. Different prediction steps represent that the value of the h-th time step after the input time window is taken as the prediction target. The experimental results of the multivariate time series prediction model on the test set are shown in Tables 1-4.

[0105] Table 1 Performance of the model on the Nature-gas-1 dataset

[0106]

[0107] Table 2 Performance of the model on the Nature-gas-2 dataset

[0108]

[0109] Table 3 Performance of the model on the Solar-energy dataset

[0110]

[0111] Table 4 Performance of the model on the Electricity dataset

[0112]

[0113] It can be seen that the multivariate time series prediction model of the present application has good performance in different energy quantity and different time step prediction tasks, and can accurately predict different future energy quantities.

[0114] In addition, as the trend of energy quantity changes over time, the multivariate time series prediction model stored in the above energy prediction module needs to be continuously learned online according to the latest data. Therefore, in another embodiment of the present application, the above energy time series data prediction system further comprises an online learning module for continuously extracting newly added data from the time series stored by each energy quantity monitoring device and constructing it as a training sample of the above multivariate time series prediction model. After the training samples accumulate to a specified number or the distance from the last model training reaches a specified time, the training samples can be used to train the multivariate time series prediction model again, realizing online learning of the multivariate time series prediction model. After each training is completed, the latest trained multivariate time series prediction model can be stored in the energy prediction module for actual prediction tasks.

[0115] It should be noted that in the above energy time series data prediction system, the multivariate time series prediction model initially stored in the energy prediction module can be offline trained or directly trained online by importing historical time series data into the platform and calling the online learning module.

[0116] It should be noted that in the above energy time series data prediction system, the data online collection module is real-time working, while the input sequence extraction module and the energy amount prediction module only perform work after receiving a prediction instruction. The prediction instruction can be sent by the system itself periodically or triggered by an external device or terminal.

[0117] Therefore, based on the same inventive concept, in another embodiment of the present application, an energy time series data prediction platform is provided, which runs the energy time series data prediction system as described in the above embodiments, and provides data access and instruction input interfaces to the outside.

[0118] It should be noted that the prediction platform mentioned here can be a local server or a cloud platform.

[0119] Based on the same inventive concept, in another embodiment of the present application, an intelligent terminal device is provided, which can establish a communication connection with the energy time series data prediction platform in the above embodiments, and access the data in the platform and send task instructions to the platform through the interfaces provided by the prediction platform.

[0120] The task instructions sent by the intelligent terminal device to the platform can include prediction instructions for executing prediction tasks, retraining instructions for calling the online learning module to retrain the model, and specification instructions for specifying the time steps required for prediction, which can be designed according to actual needs.

[0121] As Figure 5As shown, the data transmission form of the above-mentioned energy time series data prediction platform, energy amount monitoring device and intelligent terminal device in the form of a cloud platform is shown. The intelligent terminal device can be any computer electronic device that can run a program, and can be selected according to the application scenario. The above-mentioned computer electronic device needs to include a memory and a processor. The memory is used to store a computer program; the processor is used to realize the function of the corresponding system module when executing the aforementioned computer program. It can be understood that the above-mentioned memory can adopt a random access memory (Random Access Memory, RAM), and can also adopt a non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk memory. At the same time, the storage medium can also be a U disk, a mobile hard disk, a magnetic disk or an optical disk, and various storage program code media. It can be understood that the above-mentioned processor can be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP) and the like; it can also be a digital signal processor (Digital Signal Processing, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. As a preferred mode of the embodiment of the present application, the above-mentioned intelligent terminal device can be selected as a smart phone, a tablet computer, a PC.

[0122] In addition, it should be noted that in the system, platform and terminal device provided by the above-mentioned embodiment, each module is a program module executed in sequence when executed, and therefore its essence is to execute a data processing flow. It can be clearly understood by those skilled in the art that, for the convenience and brevity of description, the specific working process of the above-mentioned system can refer to the corresponding process in the foregoing method embodiment, which will not be described here. In each embodiment provided in the present application, the division of steps or modules in the method and system is only a logical functional division, and another division mode can be used in actual implementation, for example, multiple modules or steps can be combined or integrated together, or a module or step can be split.

[0123] The above-described embodiments are only the preferred ones of the present application, and are not intended to limit the present application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application. Therefore, any technical solutions obtained by equivalent replacement or equivalent transformation shall fall within the protection scope of the present application.

Claims

1. An energy time series data prediction system for predicting a respective monitored energy quantity, which is an energy consumption quantity or an energy production quantity, in cooperation with a plurality of energy quantity monitoring devices, characterized in that, The utility model relates to a kind of energy consumption prediction method based on multi-element time series, including: Data online acquisition module is used to receive the real-time energy consumption data uploaded by each energy consumption monitoring device through Internet of Things, and it is stored in time series form with time stamp after being associated with monitoring device identification number; Input sequence extraction module is used to extract the latest energy consumption data sequence from the time series stored by each energy consumption monitoring device according to the received prediction instruction in a fixed length time window, and combine all sequence segments after normalizing each sequence segment to form a multi-element time series segment; Energy consumption prediction module is used to input the multi-element time series segment generated by the input sequence extraction module into the pre-trained multi-element time series prediction model according to the received prediction instruction to obtain the energy consumption prediction value monitored by each energy consumption monitoring device at a specified time step in the future. The multi-element time series prediction model includes a segmented graph convolution module, a time dilation convolution module and a time information enhancement module, and the input of the model is a multi-element time series segment;The segmented graph convolution module segments the input multi-element time series segment according to the time dimension, constructs a segmented adjacency matrix according to the cosine similarity between the sequences in the segment, and then performs multi-layer segmented graph convolution to extract the correlation between the sequences to obtain a first output signal;The time dilation convolution module inputs the first output signal output by the segmented graph convolution module into a multi-layer stacked threshold control unit with dilation convolution to further extract the features within each sequence and reduce the dimension to obtain a second output signal;The time information enhancement module inputs the input multi-element time series segment into the LSTM network, groups the signals of each hidden step according to time, applies attention mechanism to each group to extract segmented context vectors, then applies attention mechanism on the segmented context vectors to generate time context vectors, finally concatenates the time context vectors with the last signal of the LSTM network hidden layer, and passes the concatenated result through a fully connected layer to obtain a third output signal;Finally, the second output signal and the third output signal are added, and the final prediction output of the multi-element time series prediction model is obtained after inverse normalization.

2. The energy time series data forecasting system of claim 1, wherein, It also includes an online learning module for continuously extracting newly added data from the time series stored by each energy consumption monitoring device and constructing it as a training sample for the multi-element time series prediction model, and retraining the multi-element time series prediction model with these training samples when the training samples accumulate to a specified number or reach a specified time from the last model training to realize online learning of the multi-element time series prediction model.

3. The energy time series data forecasting system of claim 1, wherein, In the input sequence extraction module, the sequence segment is normalized by taking the maximum value in the entire time series where the sequence segment is located as the scaling value.

4. The energy time series data forecasting system of claim 1, wherein, In the segmented graph convolution module, 2-layer graph convolution is used to extract the correlation between sequences;In the time dilation convolution module, 5-layer threshold control unit with dilation convolution is used to extract the features within each sequence.

5. The energy time series data forecasting system of claim 1, wherein, The energy amount is a natural gas flow monitored by a natural gas flow metering device, or is an electricity consumption monitored by an electricity metering device, or is a photovoltaic power generation monitored by a photovoltaic power station.

6. An energy time series data forecasting platform, characterized in that, The prediction platform runs the energy time series data prediction system as claimed in any one of claims 1-5, and provides data access and instruction input interfaces to the outside.

7. The energy time series data forecasting platform of claim 6, wherein, The prediction platform is a local server or a cloud platform.

Images