End-side cloud framework-oriented long-time-sequence prediction multistage optimization system and method thereof

Abstract

The invention discloses a long-time-sequence prediction multistage optimization system and method for an end-side cloud framework, the multistage optimization system is embedded into an end-side cloud cooperation framework, and the end-side cloud cooperation framework comprises an end side model, an edge side model and a cloud side model which are sequentially arranged from bottom to top; each of the end side model, the edge side model and the cloud side model is composed of a data set integration unit, a feature relation extraction unit, a corresponding model distillation training unit and a prediction task conversion unit; the end-side cloud collaboration framework is trained autonomously through a cloud side model and guides training of a lower-layer model step by step from top to bottom; the lower-layer model receives training guidance from the upper-layer model and a part of same-level models; according to the method, the prediction precision is improved, the data privacy is enhanced, the use of computing resources of edge equipment is optimized, and the problems of model compression and knowledge transmission are solved.

Description

Technical Field

[0001] The present invention belongs to the field of computer networks, and in particular to a long-term prediction multi-level optimization system and method for an edge-to-edge cloud framework. Background Art

[0002] In recent years, the Internet of Things (IoT) technology has made significant progress, and more and more devices and sensors have been deployed in application scenarios such as smart homes, industrial automation, smart cities, and telemedicine. Figure 1 As shown in the figure. These edge sensors continuously collect and generate a large amount of time series data, such as environmental monitoring data, equipment operation status data, and user behavior data. With the rapid development of the Internet of Things and edge computing technologies, the generation of massive edge time series data has led to an increasing demand for efficient response processing. Due to network bandwidth limitations and data privacy issues, traditional cloud-based centralized processing methods are difficult to meet the requirements of high real-time request responses and security of large-scale data transmission. Therefore, deploying efficient long-series prediction models on edge devices has become an important solution. However, the computing power of edge devices is limited. How to adapt to the resource constraints of the edge computing environment while ensuring model accuracy is a key challenge.

[0003] With the rapid growth of data volume, traditional centralized data processing and analysis methods face severe challenges. Transmitting all data to centralized cloud servers for processing not only increases the burden on network bandwidth, but also increases latency, affecting the performance of applications with high real-time requirements. In addition, the process of transmitting large amounts of sensitive data to the cloud also brings serious data privacy and security issues.

[0004] In the edge-cloud collaborative framework, deploying long-term prediction models near edge devices can effectively reduce dependence on network resources and significantly improve service quality and response speed. However, due to the limited computing power of edge devices, it is necessary to find an optimal balance between the number of model parameters and prediction accuracy to ensure the efficient operation of the model on edge devices. Different edge devices face different service scenarios. Their physical isolation and independence of data processing bring another challenge, that is, how to achieve cross-device knowledge sharing to improve training effects while ensuring data privacy. Since edge devices process and store data independently, the logical isolation of deployed models increases the complexity of cross-device knowledge transfer and sharing. Therefore, developing a method that can effectively share knowledge between edge devices, reduce computing power requirements, and do not lose model prediction accuracy has become the key to improving the prediction capabilities of edge devices and optimizing the edge-cloud collaborative framework.

[0005] As a common model compression method, knowledge distillation technology is an important tool for optimizing deep learning models. At the same time, the ability to transfer the intrinsic knowledge of neural networks between models is also of great concern. It enables models to adapt to resource-constrained computing environments and improves the practicality and reliability of models. In edge computing, smart devices, and other scenarios with high requirements for model efficiency and performance, knowledge distillation provides an effective model compression and optimization solution.

[0006] At the same time, edge devices that are close to each other in the spatial dimension may have a certain degree of intrinsic correlation. The time series data generated in the same time interval has a certain intrinsic connection, which can assist model training to a certain extent. In the field of images, convolution operations, as a core technology, play a vital role in feature extraction. The convolution operation is applied to the time series data generated by different edge devices, the intrinsic relationship between different device scenarios is extracted, and the correlation between data features is strengthened, thereby improving the prediction accuracy and accelerating convergence. Therefore, the background technology mainly involved in the present invention includes the following aspects.

[0007] Edge computing As the demand for data processing increases, edge computing technology has emerged. It provides artificial intelligence services for the rapidly growing terminal devices and data, improving business stability. Edge computing is close to the data source (such as smart terminals), stores and processes data at the edge of the network, has location awareness, and provides proximal services to users. In terms of data processing, edge computing is faster, real-time and secure. In addition, it can solve the problem of excessive energy consumption in cloud computing, reduce costs, and alleviate network bandwidth pressure. Edge computing has been applied to many fields such as production, energy, smart home, and transportation.

[0008] Knowledge distillation is a model compression and acceleration technology that enables efficient deployment of deep learning models on resource-limited devices by transferring knowledge from a large teacher model to a small student model. It transfers the knowledge of the teacher model to the student model, thereby reducing the number of parameters and storage requirements of the student model, reducing the complexity of the model, and enabling the student model to reason faster on resource-limited devices. By extracting knowledge from the teacher model, knowledge distillation allows small models to benefit from the features and representations learned by large models on large-scale data.

[0009] Convolutional feature extraction Convolutional neural networks (CNNs) are commonly used for image recognition and processing tasks. At its core is the convolution operation, which extracts local features by sliding a convolution kernel (a small piece of weight) over the input data, effectively capturing the spatial structure information of the image. The convolution operation slides the convolution kernel over the input data to extract features and capture the local structure of the space. The network applies the convolution kernel at different locations to learn and recognize various patterns in the image, such as edges, textures, and shapes. This feature extraction process enables the CNN to learn higher-level abstract representations during training, so that it can distinguish between different objects or image categories. Summary of the invention

[0010] In view of the difficulties in the existing technology, the present invention provides a multi-level optimization system and method for long-term time series prediction for the edge-end cloud framework; the invention uses edge-end cloud collaborative training and knowledge distillation technology, combined with inline feature extraction and cross-model data rectification algorithms, to achieve adaptation to multiple tasks and resources. Its purpose is to reduce the computing power dependence of the edge-end long-term time series prediction model, improve reasoning accuracy, and protect data privacy.

[0011] In order to solve the problems of the prior art, the present invention adopts the following technical solution: (Fill in after confirmation)

[0012] A multi-level optimization system for long-term time series prediction for a device-edge-cloud framework, wherein the multi-level optimization system is embedded in a device-edge-cloud collaborative framework, and comprises: a device-side model, an edge-side model, and a cloud-side model arranged in sequence from bottom to top; the device-side model, the edge-side model, and the cloud-side model are all composed of a data set aggregation unit, a feature relationship extraction unit, a corresponding model distillation training unit, and a prediction task conversion unit; wherein:

[0013] The multi-level optimization system is trained autonomously through the cloud-side model, and guides the training of the lower-level models step by step from top to bottom; the lower-level models receive training guidance from the upper-level models and some models of the same level;

[0014] The data set aggregation unit sends the local data set to the upper model that manages it through the lower model; the upper model receives the data set of the lower model and aggregates it into a new local data set, and continues to send it to the upper model to complete the data set aggregation;

[0015] The feature relationship extraction unit performs prediction task relationship purification and information aggregation by extracting deep-level features of time series feature correlation;

[0016] The corresponding model distillation training unit extracts accurate features from the time series of models at different levels as prediction tasks for knowledge sharing training;

[0017] The prediction task conversion unit converts the time series in the prediction tasks of different level models to achieve prediction task connection.

[0018] Furthermore, the device-edge-cloud collaborative framework is trained autonomously through the cloud-side model, and guides the training process of the lower-layer models step by step from top to bottom, including:

[0019] For the nth layer model of the system n and time series The training process is as follows:

[0020]

[0021]

[0022] in: Model n The prediction results, Model n+1 The prediction results carry higher-level abstract knowledge; CP(·) is the feature relationship extraction unit, RH(·) is the prediction task conversion unit, and Train(·) is the model training.

[0023] Furthermore, the multi-level optimization system for long-term time series prediction for the edge-cloud framework is characterized in that: the data set aggregation unit sends the local data set to the upper-level model that manages it through the lower-level model; the upper-level model receives the data set of the lower-level model and aggregates it into a new local data set, and continues to send it to the upper-level model to complete the data set aggregation process, including:

[0024] Align the time dimension and treat the time series of a device on the end as a one-dimensional feature. Aggregate different device data sets in the feature dimension.

[0025] Through the one-way flow of data sets, data is aggregated from the client side to the cloud side, and each node has only the data set necessary for local model training. At the same time, since subsequent operations only involve the flow of knowledge between models, the models connected to the framework cannot obtain data at the same or higher layers, ensuring data privacy under the entire framework.

[0026] For the edge model and cloud side model, let Represents the dataset on the i-th device at the n-th layer in the edge-cloud collaborative framework. This dataset is used as the input dataset of the i-th model at the n-th layer of the multi-level optimization system. Since the device is responsible for managing j devices on the n-1th layer, The following conditions should be met:

[0027]

[0028] Where k is the starting index of the managed device; Represents the time series data set collected and generated by the i-th end-side device. The devices at this layer only use local data for training, without data aggregation, and only send it to the upper layer.

[0029] Furthermore, the feature relationship extraction unit performs prediction task relationship purification and information aggregation process by extracting deep-level features of time series feature correlation, including:

[0030] By stacking multiple convolutions, the relationship between the features of models at different levels is purified; at the same time, the step size of the convolution is selected to be the same as the size of the convolution kernel;

[0031] For a one-dimensional convolution kernel K with length and step size a, 1 ,…,k a ), and the input time series The process of feature relationship extraction is as follows:

[0032]

[0033] in, is the first Features, is the first layer feature relationship extraction sequence features, t is the layer number of the layers other than the first layer in the feature relationship extraction, is the enhanced feature relationship obtained by the feature relationship extraction unit, It is the data used for model training later. Further, the corresponding model distillation training unit extracts accurate features from the time series of models at different levels as prediction tasks to perform knowledge sharing training process, including:

[0034] For the n-th level model Model n , the knowledge distillation process and loss value calculation process are as follows:

[0035]

[0036] Loss KD =α×Loss teacher +(1-α)×Loss turth

[0037] in: Model n+1 predictions, Model n The predicted value during knowledge distillation, is the true value, and α∈[0,1] is a hyperparameter that controls the distillation ratio.

[0038] Furthermore, the prediction task conversion unit converts the time series in the prediction tasks of the different level models to realize the prediction task connection process, including: the input layer converts the input data into the format of the embedding layer input; the embedding layer cuts or expands the input data in order according to the recorded position; the output layer converts the output data of the embedding layer into the format of model training;

[0039] When the lower model passes the data to the upper model for prediction, the original sequence It will be moved to the upper time series by The corresponding positions are expanded in order to obtain the time series

[0040]

[0041] When the upper model passes the prediction results to the lower model for learning, the prediction sequence The time series will be cut according to the recorded position using the following formula:

[0042]

[0043] Beneficial Effects

[0044] Compared with the traditional technical solution, the beneficial effects brought by the present invention are:

[0045] The present invention improves the prediction accuracy of the edge model while avoiding the large-scale dissemination of data in the network, which may cause privacy data leakage. In view of the problems existing in traditional model compression methods when compressing cloud models, such as the excessively high model compression rate leading to the dilution of prediction knowledge and the low adaptability of global knowledge to edge device scenarios, the present invention proposes feature intrinsic correlation enhancement technology and a knowledge sharing framework, which enables the model to be trained for the corresponding edge scenarios, and at the same time obtains the prediction knowledge of other similar device models under the edge-cloud collaborative framework to assist in training, thereby improving the model accuracy.

[0046] The present invention realizes efficient prediction model deployment on edge devices through multi-level knowledge sharing technology and feature intrinsic correlation enhancement technology. First, the knowledge distillation method is adopted to share abstract knowledge from different levels in the end-edge-cloud collaborative framework to ensure that the model obtains accurate features of the time series. Secondly, a "feature relationship extraction unit" based on convolution operation is proposed to perform relationship extraction and information aggregation at the feature level to explore the deep correlation between time series features. Finally, the "prediction task conversion unit" is used to promote the synchronization of knowledge features between different purification mechanisms, thereby enhancing the stability and robustness of the framework. This invention not only improves prediction accuracy and enhances data privacy, but also optimizes the use of computing resources of edge devices and solves the problems of model compression and knowledge transfer. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of an application scenario of a long-term prediction multi-level optimization system and method thereof for an end-edge-cloud framework of the present invention;

[0048] Figure 2 It is a schematic diagram of the structure of a long-term prediction multi-level optimization system and method for an end-edge-cloud framework of the present invention;

[0049] Figure 3 It is a flow chart of a feature relationship extraction unit in a long-term prediction multi-level optimization system for an end-edge cloud framework and a method structure thereof of the present invention. DETAILED DESCRIPTION

[0050] The following is combined with Figure 1 to Figure 3 The present invention is described as follows:

[0051] The present invention provides a long-time series prediction multi-level optimization system and method for an end-edge-cloud framework. First, inspired by the knowledge distillation technology, the present invention proposes a "multi-level optimization system" to share prediction knowledge from top to bottom in the end-edge-cloud collaborative framework, so that the system can obtain prediction knowledge of time series from different levels while focusing on the information and prediction tasks at the current level. Secondly, inspired by the convolution operation, a deep feature structure "feature relationship extraction unit" based on the correlation of time series features is proposed to perform relationship extraction and information aggregation at the feature level to help the model mine the correlation between time series features. Finally, in order to promote the synchronization of knowledge feature dimensions between different purification mechanisms, the present invention proposes a "prediction task conversion unit" as a pipeline connecting different distillation models, so that models at different levels can be freely connected, improving the stability and robustness of the framework.

[0052] (1) Introduction and Background

[0053] Problem definition command Represents the input data of the nth layer model of the multi-level optimization system. The length of this time series is L and the number of features is M n It consists of multiple lower-level time series The task of the present invention is to train a simpler model at the lower level of the multi-level optimization system, that is, the model close to the end side. n , and make it predict better results on the corresponding end-side device. Therefore, the goal of the present invention is to minimize the loss function To make the model's prediction value Closer to the true value.

[0054] Dataset Structure Represents the data set on the i-th device in the n-th layer of the edge-cloud collaborative framework. This data set serves as the input data of the i-th model in the n-th layer of the multi-level optimization system. Since the device is responsible for managing j devices on layer n-1, The following conditions should be met:

[0055]

[0056] Where k is the starting index of the managed device. In particular, Represents the time series data set collected and generated by the i-th end-side device.

[0057] (2) Framework Introduction

[0058] like Figure 2 As shown, the end-edge-cloud collaborative framework of the present invention includes: an end-side model, an edge-side model, and a cloud-side model arranged in sequence from bottom to top; the end-edge-cloud collaborative framework adopts a multi-level optimization system, in which: the cloud-side model is trained autonomously, and then guides the training of lower-level models step by step from top to bottom. The lower-level models receive training guidance from higher-level models and some models of the same level. The end-side model, edge-side model, and cloud-side model are all composed of a data set aggregation unit, a feature relationship extraction unit, a corresponding model distillation training unit, and a prediction task conversion unit. Among them:

[0059] The data set aggregation unit: The lower-level device needs to send the local data set to the higher-level device that manages it. The higher-level device receives the data sets of multiple lower-level devices and aggregates them into a new local data set, which is then sent to a higher level to complete the data set aggregation.

[0060] The corresponding model distillation training unit: through knowledge distillation, uses the deeper prediction knowledge in the upper model to guide the training of the lower model, and shares the prediction knowledge contained in it among several lower models guided by the same upper model.

[0061] The feature relationship extraction unit: before model training, strengthens the intrinsic relationship between data set features (i.e., time series data of different end-side devices).

[0062] The prediction task conversion unit: connects each model during the corresponding model distillation training process to ensure smooth data flow. The multi-level optimization system (multi-level framework, corresponding model distillation training), feature relationship extraction unit and prediction task conversion unit are the main parts of the present invention, which will be explained in detail later.

[0063] (A) The edge-cloud collaborative framework uses a multi-level optimization system

[0064] The multi-level optimization system is designed to extract abstract knowledge from time series data at different levels to guide the training of end-side models. The present invention proposes a multi-level framework in which models at each level can extract abstract knowledge of different depths from time series and give priority to completing the prediction tasks at the current level. The abstract knowledge of different depths extracted by different models from different levels of scenes can reveal the growth pattern of time series in the corresponding device scenarios. Similar to predictions based on trends, periodicity, and volatility, the development laws in similar scenarios at the same level and deep scenarios at a higher level can help the system accurately predict time series using abstract knowledge at different levels. Specifically, low-level models are trained using real data and abstract knowledge generated by predictions from high-level models. For the nth layer Model n and time series The training process is as follows:

[0065]

[0066] in Model n The prediction results, Model n+1 The prediction results carry higher-level abstract knowledge. CP(·) is the feature relationship extraction unit, RH(·) is the prediction task conversion unit, and Train(·) is the model training.

[0067] Corresponding model distillation training unit For the training of a single model, the present invention adopts knowledge distillation technology to transfer prediction knowledge from high-level models to low-level models to improve the prediction ability of low-level models with simpler structures. Since the larger the interval between levels, the smaller the influence of high-level abstract knowledge on low-level, only the prediction results of the higher level are used for refinement. This method dilutes the abstract knowledge of the high level and avoids its excessive influence on the prediction of the low-level model. For the n-th level model Model n , the knowledge distillation process and loss value calculation process are as follows:

[0068]

[0069] Loss KD =α×Loss teacher +(1-α)×Loss turth

[0070] in: Model n+1 predictions, Model n The predicted value during knowledge distillation, is the true value, and α∈[0,1] is a hyperparameter that controls the distillation ratio.

[0071] (B) Feature Relationship Extraction Unit

[0072] The Softmax method in traditional knowledge distillation affects the trend and volatility of time series, making it difficult for the student model to fit the target sequence, thus hindering knowledge transfer. Inspired by convolutional neural networks, this paper proposes a dataset-level time series feature relationship extraction unit (such as Figure 3 As shown in Figure 2, the Softmax method is used to replace the Softmax method. This method aims to purify the intrinsic relationships and structural features between different features in the time series. The purified feature relationships are added to the original sequence in the form of catalysts. By expanding the perception field and strengthening the feature relationships, the model can better fit the characteristics of the original sequence.

[0073] In this method, each convolution purifies the relationship and structural commonality between time series features within a certain range. By superimposing multiple convolutions, the relationship between features at different levels can be purified. At the same time, the step size of the convolution is selected to be the same as the size of the convolution kernel to avoid the relationship granularity after purification being too fine, resulting in an explosion in the number of feature relationships. For a one-dimensional convolution kernel K=(k 1 ,…,k a ), and the input time series The process of feature relationship extraction is as follows:

[0074]

[0075] When the convolution kernel size and step size are selected to be larger, fewer feature relationships are purified in a single iteration. Therefore, the model requires fewer iterations, but the level of detail and purification is lower. Conversely, using a smaller kernel size and step size can achieve higher detail granularity, but at a slower speed. The characteristic number can be derived from the following equation:

[0076]

[0077] Among them, M n For the input time series The number of features, Output time series for feature relationship extraction unit The number of features.

[0078] (C) Prediction Task Switching Unit

[0079] In order to solve the problem of feature quantity differences caused by different scales of data sets at different levels, the present invention introduces a prediction task conversion unit in the multi-level distillation framework. in and Zoom outThe mechanism solves the challenges of different dimensions of time series at different levels. The mechanism consists of an embedding layer with position recording function, which is implemented by positioning the expansion / cropping time series. The introduction of the prediction task conversion unit improves the stability and robustness of the multi-level distillation framework. Its specific working principle is as follows:

[0080]

[0081] Among them, for Zoom out (·) function, which converts the original sequence Expand in order and move it to the upper time series Corresponding location. For Zoom in (·)$ function, which converts the original sequence Crop by recorded position to get time series

[0082] For the Trans(·) function, it first converts the time series into a format that can be used as an input to the embedding layer, and then converts it into a format that can be used for model training. This method ensures that the lower-level model can obtain more abstract knowledge that is more suitable for the current task from the upper-level model, under the premise that the upper-level dataset is aggregated from the lower-level dataset.

[0083] Although the present invention has been described above, the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many modifications without departing from the purpose of the present invention, which are all within the protection of the present invention.

Claims

1. A long-term prediction multi-level optimization system for an edge-cloud framework, characterized by: The multi-level optimization system is embedded in the device-edge-cloud collaborative framework, and the multi-level optimization system includes: a device-side model, an edge-side model, and a cloud-side model arranged in sequence from bottom to top; the device-side model, the edge-side model, and the cloud-side model are all composed of a data set aggregation unit, a feature relationship extraction unit, a corresponding model distillation training unit, and a prediction task conversion unit; wherein: The multi-level optimization system is trained autonomously through the cloud-side model, and guides the training of the lower-level models step by step from top to bottom; the lower-level models receive training guidance from the upper-level models and some models of the same level; The data set aggregation unit sends the local data set to the upper model that manages it through the lower model; the upper model receives the data set of the lower model and aggregates it into a new local data set, and continues to send it to the upper model to complete the data set aggregation; The feature relationship extraction unit performs prediction task relationship purification and information aggregation by extracting deep-level features of time series feature correlation; The corresponding model distillation training unit extracts accurate features from the time series of models at different levels as prediction tasks for knowledge sharing training; The prediction task conversion unit converts the time series in the prediction tasks of different level models to achieve prediction task connection.

2. According to claim 1, a long time series prediction multi-level optimization system for the edge-cloud framework is characterized by: The device-edge-cloud collaborative framework is trained autonomously by the cloud-side model and guides the training process of the lower-level models step by step from top to bottom, including: For the nth layer model of the system n and time series The training process is as follows: in: Model n The prediction results, Model n+1 The prediction results carry higher-level abstract knowledge; CP(·) is the feature relationship extraction unit, RH(·) is the prediction task conversion unit, and Train(·) is the model training.

3. According to claim 1, a long time series prediction multi-level optimization system for the edge-cloud framework is characterized by: The data set aggregation unit sends the local data set to the upper model that manages it through the lower model; the upper model receives the data set of the lower model and aggregates it into a new local data set, and continues to send it to the upper model to complete the data set aggregation process, including: Align the time dimension and treat the time series of a device on the end as a one-dimensional feature. For data sets of different devices, aggregate them in the feature dimension. Through the one-way flow of data sets, data is aggregated from the client side to the cloud side, and each node has only the data set necessary for local model training. At the same time, since subsequent operations only involve the flow of knowledge between models, the models connected to the framework cannot obtain data at the same or higher layers, ensuring data privacy under the entire framework. For the edge model and cloud side model, let Represents the dataset on the i-th device at the n-th layer in the edge-cloud collaborative framework. This dataset is used as the input dataset of the i-th model at the n-th layer of the multi-level optimization system. Since the device is responsible for managing j devices on the n-1th layer, The following conditions should be met: Where k is the starting index of the managed device; Represents the time series data set collected and generated by the i-th end-side device. The devices at this layer only use local data for training, without data aggregation, and only send it to the upper layer.

4. According to claim 1, a long time series prediction multi-level optimization system for an edge-to-edge cloud framework is characterized by: The feature relationship extraction unit performs prediction task relationship purification and information aggregation process by extracting deep-level features of time series feature correlation, including: By stacking multiple convolutions, the relationship between the features of models at different levels is purified; at the same time, the step size of the convolution is selected to be the same as the size of the convolution kernel; For a one-dimensional convolution kernel K=(k1,…,k a ), and the input time series The process of feature relationship extraction is as follows: in, is the first Features, is the first layer feature relationship extraction sequence features, t is the layer number of the layers other than the first layer in the feature relationship extraction, is the enhanced feature relationship obtained by the feature relationship extraction unit, This is the data used for model training later.

5. According to the long-term prediction multi-level optimization system for the edge-cloud framework of claim 1, Features: The corresponding model distillation training unit extracts accurate features from the time series of models at different levels as prediction tasks to perform a knowledge sharing training process, including: For the n-th level model Model n , the knowledge distillation process and loss value calculation process are as follows: Loss KD =α×Loss teacher +(1-a)×Loss turth in: Model n+1 predictions, Model n The predicted value during knowledge distillation, is the true value, and α∈[0,1] is a hyperparameter that controls the distillation ratio.

6. According to claim 1, a long time series prediction multi-level optimization system for the edge-cloud framework is characterized by: The prediction task conversion unit converts the time series in the prediction tasks of the different level models to realize the prediction task connection process, including: the input layer converts the input data into the format of the embedding layer input; the embedding layer cuts or expands the input data in order according to the recorded position; the output layer converts the output data of the embedding layer into the format of model training; When the lower model passes the data to the upper model for prediction, the original sequence It will be moved to the upper time series by The corresponding positions are expanded in order to obtain the time series When the upper model passes the prediction results to the lower model for learning, the prediction sequence The time series will be cut according to the recorded position using the following formula:

7. A multi-level optimization method for long-term time series prediction for an edge-cloud framework, characterized by: The predictive multi-level optimization method is performed based on the system of any one of claims 1-6.

Images