Prediction platform for predicting time series data of energy amount
The energy time series data prediction system addresses the lack of universal models by using a multivariate time series prediction model to accurately forecast energy consumption and generation across different industries, enhancing energy sector development with real-time data management and adaptive learning.
Patent Information
- Application Number
- JP2024114463
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-03-23
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-03-18
AI Technical Summary
There is no common platform or model with universality for predicting energy time series data across different energy industries, leading to duplication of construction efforts and limiting the development of the energy sector.
An energy time series data prediction system utilizing a multivariate time series prediction model that includes a data online collection module, input sequence extraction module, and energy amount prediction module, employing segment graph convolution, time dilation convolution, and temporal information reinforcement to predict energy consumption or generation.
The system provides accurate and versatile predictions for various energy types, supporting real-time data publication and visualization, and enables online learning to adapt to changing conditions.
Smart Images

Figure 2024156234000127 
Figure 2024156234000128 
Figure 2024156234000129
Abstract
Description
[Technical field]
[0001] The present invention relates to the field of time series data forecasting, and more particularly to an energy time series data forecasting system, platform and smart terminal device. [Background technology]
[0002] In recent years, with the development of Internet of Things (IoT) technology, various sensors have the ability to upload in time or in real time, generating a large amount of multivariate time series data composed of different sensors. The sensors are monitoring devices, and the series data generated by monitoring is highly related to the monitored energy devices behind them. For energy amount monitoring devices such as natural gas meters, solar power plants, and power meters, they can generate time series data related to energy consumption or energy generation through monitoring, and there is a demand for predicting future trends for all of these energy amount related data.
[0003] In the energy field, sensors do not always have a natural physical relationship, and the usage rules of monitored devices may change dynamically and have a high degree of time correlation, so the prediction model needs to model the characteristics of the data to improve the prediction accuracy. Realizing the prediction of energy data can, on the one hand, analyze the usage or production rules of energy devices, and, on the other hand, provide reference for energy procurement for energy companies, thereby reducing costs and generating economic benefits.
[0004] Energy time series data is widely available in different energy industries, such as natural gas, electricity, geothermal energy, etc. However, there is currently no common platform and model for predicting different energy quantities that is universal for different prediction targets. Building prediction platforms and models for different industries will, on the one hand, cause duplication of construction, and on the other hand, limit the development of the energy industry, where some technical capabilities are weak.
[0005] The objective of the present invention is to solve the problem in the prior art of lacking a common platform and model that is universal for different prediction objects, and to provide an energy time series data prediction system, platform and smart terminal device.
[0006] The specific technical solutions adopted in the present invention are as follows: In a first aspect, the present invention provides an energy time series data prediction system. The energy time series data prediction system is used to predict the monitored energy amounts in cooperation with a plurality of energy amount monitoring devices. The energy amounts are energy consumption amounts or energy generation amounts. The energy time series data prediction system includes a data online collection module, an input sequence extraction module, and an energy amount prediction module. The data online collection module is used to receive the real-time energy quantity data uploaded by each of the above energy quantity monitoring devices via the Internet of Things, and store it in a time-stamped time series format after associating it with the monitoring device identification number. The input sequence extraction module is used to extract the latest segment of energy quantity data sequence from the time series stored in each energy quantity monitoring device based on the received prediction command based on a fixed-length time window, and combine all the extracted sequence segments into a multivariate time series segment after normalizing them respectively. The energy amount prediction module inputs the multivariate time series segments generated by the input series extraction module into a pre-trained multivariate time series prediction model based on the received prediction command, and is used to obtain a predicted value of the energy amount monitored by each energy amount monitoring device at a specified time step in the future.
[0007] In a preferred embodiment of the first aspect, the multivariate time series prediction model may include a segment graph convolution module, a time dilation convolution module, and a time information augmentation module, and the input of the model may be a multivariate time series segment. The segment graph convolution module segments the input multivariate time series segment according to a time dimension, constructs a segment adjacency matrix based on the cosine similarity between the series in the segment, and further performs multi-layer segment graph convolution to extract correlation between the series and obtain a first output signal. The time dilation convolution module inputs the first output signal output from the segment graph convolution module to a gated linear unit with multi-layer stacked dilation convolution, and further extracts features in each series and performs dimensional reduction to obtain a second output signal. The above-mentioned time information augmentation module inputs the input multivariate time series segments into the LSTM network, groups the signals of each hidden step of the LSTM network according to time, and applies an attention mechanism to each group to extract a segment context vector, then applies an attention mechanism to the segment context vector to generate a time context vector, and finally splices the time context vector with the last signal of the hidden layer of the LSTM network, and the splicing result passes through one fully connected layer to obtain a third output signal. Finally, the second output signal and the third output signal are added and inversely normalized to obtain the final forecast output of the multivariate time series forecasting model.
[0008] As a preferred example of the first aspect, the present invention may further include an online learning module, which continuously extracts newly added data from the time series stored in each energy amount monitoring device and builds it into training samples of the multivariate time series prediction model, and reuses these training samples to train the multivariate time series prediction model after a specified number of training samples are accumulated or a specified time has passed since the previous model training, thereby realizing online learning of the multivariate time series prediction model.
[0009] As a preferred example 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 may be normalized as a scale value.
[0010] In a preferred embodiment of the first aspect, the segment graph convolution module employs two-layer graph convolution to extract correlations between sequences, and the time dilation convolution module may employ a gated linear unit with five layers of dilated convolution to extract features within each sequence.
[0011] As a preferred example of the first aspect, the amount of energy may be a natural gas flow rate obtained by monitoring with a natural gas flow metering device, an amount of electricity consumption obtained by monitoring with a power meter device, or an amount of solar power generation obtained by monitoring with a solar power plant.
[0012] In a second aspect, the present invention provides an energy time series data forecasting platform, in which the energy time series data forecasting system according to any one of the first aspects operates, and the energy time series data forecasting platform provides a data access and command input interface to the outside.
[0013] As a preferred example of the second aspect, the prediction platform may be a local server or a cloud platform.
[0014] In a third aspect, the present invention provides a smart terminal device, which can establish a communication connection with the energy time series data prediction platform of the second aspect, access data in the platform through an interface provided by the energy time series data prediction platform, and send task commands to the platform.
[0015] As a preferred example of the third aspect, the task instructions sent by the smart terminal device to the platform may include a prediction instruction to execute a prediction task, a retraining instruction to call an online learning module to re-train the model, and a specification instruction to specify a time step that needs to be predicted.
[0016] The present invention has the following beneficial effects over the prior art: 1) The present invention provides an energy time series data forecasting system for multivariate time series forecasting of energy quantity, in which a data online collection module collects real-time energy quantity data and constructs a time series, and then when forecasting is required, an input sequence extraction module extracts the latest multivariate time series segment, and then forecasts using a pre-trained multivariate time series forecasting model. The energy time series data forecasting system is universal to all different types of energy quantity forecasting. 2) In the system of the present invention, a multivariate time series forecasting model based on deep learning is designed. The model constructs a segment graph convolutional layer to make different time series into nodes in a graph, learns related information of adjacent nodes from the graph structure based on sequence similarity, extracts internal information of the time series through a time dilation convolutional layer, and combines the output of a time information augmentation module based on a two-stage attention mechanism with the output of the model. Prediction experiments on a multivariate time series dataset of actual energy amounts have shown that the model can successfully complete predictions of different energy amounts and different time steps, and can accurately predict the energy amounts of different future times. 3) The present invention can provide a universal platform for energy data such as electricity, natural gas, etc., and can provide external forecasting services, and the platform can be linked with corresponding smart terminals to realize real-time disclosure, query and visualization of forecast tasks. [Brief description of the drawings]
[0017] [Figure 1] 1 is a schematic diagram of a module configuration of an energy time-series data forecasting system according to an embodiment of the present invention. [Diagram 2] FIG. 1 is a schematic diagram of the polymerization process of single-layer graph neural network information. [Diagram 3] FIG. 1 is a schematic diagram of a network of a multivariate time series forecasting model. [Figure 4] FIG. 11 is a schematic diagram of a module configuration of an energy time-series data forecasting system according to another embodiment of the present invention. [Diagram 5] 1 is a schematic diagram showing the relationship between an energy amount monitoring device, an energy time series data prediction platform, and a smart terminal device in an embodiment of the present invention; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0018] In order to make the above-mentioned objects, features, and advantages of the present invention more clearly comprehensible, specific embodiments of the present invention will be described in detail below in conjunction with the drawings. In order to facilitate a complete understanding of the present invention, many specific details will be described in the following description. However, the present invention 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 content of the present invention, and therefore the present invention is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined correspondingly, provided that they are not mutually inconsistent.
[0019] In describing the present invention, when an element is referred to as being "connected" to another element, it is to be understood that the element may be directly connected to the other element, or may be indirectly connected, i.e., indirectly connected because of the presence of intermediate elements. Conversely, when an element is referred to as being "directly" connected to another element, there are no intermediate elements present.
[0020] The present invention provides an energy time series data prediction system, which is used in conjunction with a plurality of energy amount monitoring devices to predict the monitored energy amounts respectively. As shown in Fig. 1, in a preferred embodiment of the present invention, the energy time series data prediction system includes a data online collection module, an input sequence extraction module and an energy amount prediction module.
[0021] The data online collection module is used to receive the real-time energy quantity data uploaded by each of the above energy quantity monitoring devices via the Internet of Things, and store it in a time-stamped time series format after associating it with the monitoring device identification number. The input sequence extraction module is used to extract the latest segment of energy quantity data sequence from the time series stored in each energy quantity monitoring device based on the received prediction command based on a fixed-length time window, and combine all the extracted sequence segments into a multivariate time series segment after normalizing them respectively. The energy amount prediction module inputs the multivariate time series segments generated by the input series extraction module into a pre-trained multivariate time series prediction model based on the received prediction command, and is used to obtain a predicted value of the energy amount monitored by each energy amount monitoring device at a specified time step in the future.
[0022] In the present invention, the energy time series data prediction system does not limit the prediction target of the energy amount, and the energy amount in the present invention may be the energy consumption amount or the energy generation amount. In the following embodiment, the energy consumption amount may be the natural gas usage amount or the power usage amount, and the energy generation amount may be the solar power generation amount. Therefore, the energy amount monitoring device required for different energy amounts is also different. For example, if the energy amount is the natural gas flow rate, it can be obtained by monitoring with a natural gas flow meter device, if the energy amount is the power usage amount, it can be obtained by monitoring with a power meter device, and if the energy amount is the solar power generation amount, it can be obtained by monitoring with a solar power plant. However, in addition to the above three energy amounts, the present invention can also be applied to other energy-related time series data such as wind power generation data, oil-air production amount data, hydroelectric power generation amount data, and thermal power generation amount data.
[0023] The data transmission technology of each energy amount monitoring device may be any Internet of Things technology that can transmit data remotely. In the embodiment of the present invention, the following types may be adopted: 1. Wi-Fi technology: Wi-Fi technology is a local area network wireless transmission technology that can provide high-speed and stable network connection. By connecting devices to a Wi-Fi network, remote control and data transmission for Internet of Things devices can be realized, and it is suitable for small-range, high-speed transmission scenarios. 2. Cellular network technology: Cellular network technology is one of the core technologies in the mobile communication field. It uses 3G or 4G networks to realize remote data transmission, and has the characteristics of wide coverage range and highly efficient and stable transmission. It can be applied to cross-area, long-distance, large-volume data transmission scenarios. 3. Zigbee technology: Zigbee technology is a wireless low-power Internet of Things technology that can provide local area network level data transmission and is applicable to scenes that require low-speed, low-power transmission, such as smart homes. 4. LoRaWAN technology: LoRaWAN technology adopts long-distance, low-power wireless communication technology, which can realize remote data transmission and monitoring in different coverage and environmental conditions such as urban and rural areas.
[0024] Of course, in addition to the above-mentioned several Internet of Things technologies, other Internet of Things technologies capable of remote data transmission can also be considered and adopted, as long as they can meet the timeliness and reliability of data transmission.
[0025] In the above-mentioned data online collection module, the real-time energy quantity data received and uploaded by each energy quantity monitoring device through the Internet of Things is recorded in the data storage module using the unique identification number of the monitoring device as a unique identification, and is available for subsequent retrieval. The data uploaded by each energy quantity monitoring device can be recorded in the same time series, and each recorded value needs to be time-stamped.
[0026] In the present invention, the function of the input sequence extraction module is to extract the input data, i.e., multivariate time series segments, required for prediction of the subsequent multivariate time series prediction model. The extraction of the multivariate time series segments can be realized by sliding a time window. The multivariate time series prediction model of the present invention is used to simultaneously predict future monitoring values of multiple energy quantity monitors, so that it is necessary to extract a series of the latest segments for each time series corresponding to each energy quantity monitor by a time window, and these series are further configured to form multivariate time series segments in parallel.
[0027] Thus, the multivariate time series prediction model is the key to realizing the prediction function of the present invention. The specific realization of the multivariate time series prediction model will be described in detail below. In the multivariate time series prediction task, one N-dimensional multivariate time series sample is generated by the time changes of N variables, and each variable can correspond to one energy amount monitoring device. In the graph neural network, the model regards each variable as a node in the graph, and the degree of association between the variables is the edge in the graph structure. The graph structure is
number
number
number
number
number
number
number
[0028] Based on the above concept, as shown in Figure 3, in an embodiment of the present invention, the multivariate time series prediction model includes one segment graph convolution module, one time dilation convolution module, and one time information enrichment module. The input of the model is the multivariate time series segment. Assuming that the input of the model is the above-mentioned multivariate time series segment, and the multivariate time series segment is represented as X, the specific realization of the three modules is as follows:
[0029] (1) The segment graph convolution module segments the input multivariate time series segments based on the time dimension, constructs a segment adjacency matrix based on the cosine similarity between the series in the segment, and then performs multi-layer segment graph convolution to extract the correlation between the series and obtain the first output signal. Specifically, the multivariate time series segment X of the input model is input to the segment graph convolution module, and the final graph convolution result is
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
[0030] (2) The time dilation convolution module inputs the first output signal from the segment graph convolution module into a gated linear unit with multi-layer stacked dilation convolution, and then extracts features in each sequence and performs dimensional reduction to obtain a second output signal.
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
[0031] (3) The temporal information augmentation module inputs the input multi-dimensional time-ordered segments into an LSTM network, groups the signals of each hidden step of the LSTM network according to time, and applies an attention mechanism to each group to extract a segment context vector. Then, applies an attention mechanism to the segment context vector to generate a temporal context vector. Finally, the temporal context vector is concatenated with the last signal of the hidden layer of the LSTM network, and the concatenated result is passed through one fully connected layer to obtain a third output signal.
[0032] Specifically, the multivariate time series segment X of the input model is input to the temporal information augmentation module, and the temporal information augmentation signal
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
[0033] (4) Finally, the second output signal and the third output signal are added and then inverse normalized to obtain the final forecast output of the multivariate time series forecasting model. Specifically, the final output of the time dilation convolution module is
number
number
number
number
number
[0034] Some hyperparameters in the above multivariate time series forecasting model can be optimized based on practice. In the embodiment of the present invention, the length W of the time window is 168, and the sliding step s of the time window is 1. The segment 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 of the time dilation convolution module is 5, the number of convolution kernels M of each layer is 24, the size C1 of the convolution kernel is 1×3, and the dilation ratios D of the convolution kernels of the 5 layers are 1, 3, 6, 12, and 24, respectively. In the time dilation convolution module, the number of convolution kernels of the normal convolution of the last layer is
number
number
[0035] It should be noted that the above multivariate time series prediction model needs to be trained before application, and the specific training method belongs to the prior art. The loss function used in training can be the mean absolute error (MAE). When specifically trained, the final forecast output of the multivariate time series prediction model is
number
number
[0036] Before training the above model, a training dataset needs to be constructed in advance, which belongs to the prior art. For ease of understanding, a brief description is given below, and the flow can be seen in the following: 1) Time series recorded on N energy quantity monitoring devices that a user needs to monitor
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
[0037] After the training data set is obtained, it can be used for model training, and the random gradient of the model can be calculated based on the obtained prediction loss L. All trainable parameters of the neural network model are updated using the AdamW optimizer with a learning rate α and a decay weight β, and the training process is repeated until the model converges. Thus, during actual prediction, a real-time multivariate time series segment is constructed by obtaining a series segment within a time window closest to the present from the time series data recorded in each energy amount monitor, and the real-time multivariate time series segment is normalized and then input into the trained multivariate time series prediction model stored in the energy prediction module to predict the recorded value corresponding to each energy amount monitor at a specified time step in the future.
[0038] In one embodiment of the present invention, to demonstrate the forecasting performance that the above multivariate time series forecasting model can achieve, it is tested on natural gas datasets (Nature-gas-1 and Nature-gas-2), solar dataset (Solar-energy) and electricity dataset (Electricity). Root Relative Squared Error (RSE) and Empirical Correlation Coefficient (CORR) are selected to evaluate the performance metrics of the model. The details of some datasets are as follows: The two natural gas datasets (Nature-gas-1 and Nature-gas-2) are both derived from a natural gas flowmeter database of a gas company. The standard state integrated flow rate of the flowmeter is resampled, interpolated, and subtracted in advance to generate a natural gas standard state flow rate series. Since different natural gas sensors have different upload frequencies, two different adoption frequency natural gas datasets are constructed to display the expressions at different sampling frequencies in the multivariate time series forecasting model. Among them, Natural-gas-1 contains 52 high-frequency natural gas flow rate series from July 1, 2020 to September 5, 2021, with a sampling frequency of 10 minutes. Natural-gas-2 contains 118 low-frequency natural gas flow rate series from March 9, 2020 to September 5, 2021, with a sampling frequency of 1H. Due to the particular nature of the natural gas industry, some industrial and commercial users temporarily suspended gas use around the time of the Chinese New Year, so the data for February 2021 in the two datasets will be deleted. The solar-energy dataset (Solar-energy) contains solar production records for 2006 from 137 solar power plants from the state of Alabama, with a sampling frequency of 10 minutes. The Electricity dataset comes from the UCI Machine Learning Dataset and contains electricity usage for 321 customers from 2012 to 2014. The sampling frequency of the original data is 15 minutes, which we convert here into hourly electricity usage data. At the same time, in order to evaluate the prediction performance at different prediction steps, this embodiment sets four sets of prediction steps of the multivariate time series prediction model, that is, experiments are carried out at h=3, 6, 12, and 24, respectively, and different prediction steps mean that the value of the hth time step after the time window is input is the prediction target. The experimental results of the test set of this multivariate time series prediction model are shown in Tables 1 to 4. Table 1 Representation of the model in the Nature-gas-1 dataset [Table 1] Table 2 Representation of the model in the Nature-gas-2 dataset [Table 2] Table 3 Representation of the model in the Solar-energy dataset [Table 3] Table 4 Representation of the model in the Electricity dataset [Table 4] In this way, the multivariate time series forecasting model of the present invention is well-represented in forecasting tasks of different energy amounts and different time steps, and can accurately forecast energy amounts at different future times.
[0039] In addition, since the trend of the energy amount changes over time, the multivariate time series prediction model stored in the energy prediction module needs to always perform online learning based on the latest data. Therefore, in another embodiment of the present invention, the above energy time series data prediction system further includes an online learning module that always extracts newly added data from the time series stored in each energy amount monitoring device and constructs it as a training sample for the multivariate time series prediction model. After a predetermined number of training samples are accumulated, or after a predetermined time has passed since the previous model training, the training samples are reused to train the multivariate time series prediction model, thereby realizing online learning of the multivariate time series prediction model. After each training is completed, the multivariate time series prediction model obtained by the latest training is stored in the energy prediction module, and the actual prediction task can be performed.
[0040] It should be noted that in the above-mentioned energy time series data forecasting system, the multivariate time series forecasting model initially stored in the energy forecasting module may be trained offline, or may be trained directly online by invoking the above-mentioned online learning module after introducing historical time series data into the platform. The method of directly training online by invoking the online learning module can further lower the hurdle of using the system, and users can directly introduce related historical time series data into the platform without having to train themselves.
[0041] In the above-mentioned energy time series data prediction system, the data online collection module operates in real time, while the input sequence extraction module and the energy amount prediction module operate after receiving a prediction command. This prediction command may be periodically sent by the system itself, or may be triggered by an external device or terminal.
[0042] Therefore, based on the same inventive idea, in another embodiment of the present invention, an energy time series data forecasting platform on which the energy time series data forecasting system described in the above embodiment operates is provided, and the forecasting platform provides a data access and command input interface to the outside.
[0043] It should be noted that the prediction platform referred to here may be a local server or a cloud platform.
[0044] Based on the same inventive idea, in another embodiment of the present invention, a smart terminal device is provided, which can establish a communication connection with the energy time-series data forecasting platform in the above embodiment, access data in the platform through an interface provided by the forecasting platform, and send task commands to the platform.
[0045] The task commands sent by the smart terminal device to the platform may include a prediction command for performing a prediction task, a retraining command for calling an online learning module to re-run model training, and a specification command for specifying the time step for which prediction is required, and may be specifically designed according to actual needs.
[0046] As shown in FIG. 5, the data transmission format of the above-mentioned energy time series data forecasting platform, energy amount monitoring equipment, and smart terminal device in one cloud platform format is shown. Here, the smart terminal device is a computer electronic device capable of executing any program, and can be specifically selected according to the application scene. The above-mentioned computer electronic device requires a memory and a processor. The memory stores a computer program. The processor is for implementing the function of the corresponding system module when executing the above-mentioned computer program. It is understood that the above-mentioned memory may use a random access memory (RAM) or may use a non-volatile memory (NVM) such as at least one magnetic disk memory. In addition, the storage medium may be various media capable of storing program code, such as a USB memory, a removable hard disk, a magnetic disk, or an optical disk. It is understood that the above-mentioned processor may be a general-purpose processor including a central processing unit (Central Processing Unit), a network processing unit (Network Processor), etc. In addition, the smart terminal device may be a smartphone, tablet, or PC, although in a preferred embodiment of the present invention, the smart terminal device may be a digital signal processing (DSP), an application specific integrated circuit (ASIC), a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component.
[0047] In addition, in the system, platform, and terminal device provided by the above-mentioned embodiment, each module corresponds to a program module that is executed sequentially at the time of execution, so that it essentially executes a data processing flow. For convenience and conciseness of explanation, those skilled in the art can clearly understand that the specific operation process of the above-mentioned system can refer to the corresponding process in the above-mentioned method embodiment, and no further reference will be made here. In the embodiment provided in this specification, the division of steps or modules in the above-mentioned method and system is only one logical functional division, and there may be other division methods when actually realized, for example, multiple modules or steps may be combined or integrated, and one module or step may be divided.
[0048] The above embodiments are merely preferred aspects of the present invention and are not intended to limit the present invention. Those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Therefore, all technical solutions obtained by equivalent replacement or equivalent transformation methods are included in the protection scope of the present invention.
Claims
1. A prediction platform for predicting time series data of an energy amount on which an energy time series data prediction system operates, comprising: The energy time series data forecasting system includes: The energy amount monitoring device is used to predict the amount of energy consumed or the amount of energy generated, which are monitored in cooperation with a plurality of energy amount monitoring devices, A data online collection module; an input sequence extraction module; an energy amount prediction module; The data online collection module is used to receive the real-time energy quantity data uploaded by each of the energy quantity monitoring devices via the Internet of Things, and store the data in a time-stamped time series format after associating the data with a monitoring device identification number; The input sequence extraction module is used to extract the latest segment of energy quantity data sequence from the time series stored in each energy quantity monitoring device according to the received prediction command based on a fixed length time window, and combine all the extracted sequence segments into a multivariate time series segment after normalizing them respectively; The energy amount prediction module inputs the multivariate time series segments generated by the input sequence extraction module into a pre-trained multivariate time series prediction model according to the received prediction command, so as to obtain a predicted value of the energy amount monitored by each energy amount monitoring device at a specified time step in the future; The multivariate time series prediction model includes a segment graph convolution module, a time dilation convolution module, and a temporal information augmentation module, and the input to the model is a multivariate time series segment; The segment graph convolution module segments the input multivariate time series segments according to a time dimension, constructs a segment adjacency matrix based on cosine similarity between sequences in the segment, and further performs multi-layer segment graph convolution to extract correlation between sequences to obtain a first output signal; The time dilation and convolution module inputs the first output signal output from the segment graph convolution module into a gated linear unit with multi-layer stacked dilation and convolution, and further extracts features in each sequence and performs dimensional reduction to obtain a second output signal; The temporal information augmentation module inputs the input multivariate time series segments into an LSTM network, groups the signals of each hidden step of the LSTM network according to time, and applies an attention mechanism to each group to extract a segment context vector, and then applies an attention mechanism to the segment context vector to generate a time context vector, and finally concatenates the time context vector with the last signal of the hidden layer of the LSTM network, and passes the concatenation result through one fully connected layer to obtain a third output signal; Finally, the second output signal and the third output signal are added and denormalized to obtain a final forecast output of the multivariate time series forecasting model. The prediction platform is characterized in that it provides an interface for data access and command input to the outside.
2. The prediction platform of claim 1 , wherein the prediction platform is a local server or a cloud platform.
3. 3. The prediction platform of claim 1, wherein the energy time series data prediction system further includes an online learning module, which continuously extracts newly added data from the time series stored in each of the energy quantity monitoring devices and builds it into training samples of the multivariate time series prediction model; and reuses these training samples to train the multivariate time series prediction model after the training samples accumulate to a specified number or reach a specified time since the previous model training, thereby realizing online learning of the multivariate time series prediction model.
4. The prediction platform according to claim 1 or 2, characterized in that, in the input sequence extraction module, the energy time series data prediction system normalizes the extracted sequence segment by using a maximum value in the entire time sequence in which the sequence segment is located as a scale value when normalizing the extracted sequence segment.
5. The prediction platform of claim 1 or 2, wherein the energy time series data prediction system employs a two-layer graph convolution in the segment graph convolution module to extract correlations between series, and uses a gated linear unit with five layers of dilated convolution in the time dilation convolution module to extract features within each series.
6. 3. The prediction platform according to claim 1 or 2, wherein the amount of energy is a natural gas flow rate obtained by monitoring with a natural gas flow meter device, an amount of electricity consumption obtained by monitoring with a power meter device, or an amount of solar power generation obtained by monitoring with a solar power plant.
7. The external device is a smart terminal device, The prediction platform according to claim 1 or 2, characterized in that a communication connection is established with the smart terminal device, and the smart terminal device accesses data in the prediction platform through an interface provided by the prediction platform, and the smart terminal device transmits task commands to the prediction platform.
8. 8. The prediction platform of claim 7, wherein the task instructions include a prediction instruction to execute a prediction task, a retraining instruction to call an online learning module to perform model training again, and a specification instruction to specify a time step that needs to be predicted.
