System and method for time series prediction and analysis using large language models with a hierarchical context

Abstract

The invention presents a system and method for time series prediction using large language models (LLMs) with time-series at different scale. By leveraging LLMs operating at multiple time scales, coarser-level context, generated by alternate LLMs, machine learning models, or state machines, guides predictions. These predictions focus on network key performance indicators (KPIs) such as physical resource block (PRB) utilization, network throughput, and latency, with proactive forecasting achieved through multiple LLM blocks handling context at different time intervals. The system synthesizes multi-scale data using random processes or time-series waveforms, and integrates live network data to enhance prediction accuracy, especially in detecting anomalies. An optimization algorithm processes the predicted time series data to enable proactive, automatic network reconfiguration. The invention extends to various applications, including agriculture, cloud data centres, financial markets, smart city traffic, e-commerce, healthcare, and energy conservation. The hierarchical model captures both general trends and intricate patterns, improving time series prediction accuracy and enabling better, more proactive decision-making across industries.

Description

PJN105192SYSTEM AND METHOD FOR TIME SERIES PREDICTION AND ANALYSIS USING LARGE LANGUAGE MODELS WITH A HIERARCHIC AL CONTEXTTECHNICAL FIELD

[0001] The present disclosure relates to the field of Large Language Models (LLMs). More specifically, the present disclosure pertains to a time series prediction and analysis using LLMs with time-series at different scale.BACKGROUND OF THE INVENTION

[0002] Time series forecasting plays a crucial role in various domains, ranging from finance and economics to weather forecasting and demand planning. Traditionally, statistical models and machine learning algorithms have been employed to tackle these forecasting tasks. However, with the advent of Large Language Models (LLMs) like GPT-3.5, an exciting new tool is at our disposal.

[0003] Recently, LLMs have been widely used, achieving promising performance across various domains, such as health management, customer analysis, and text feature mining. Time series forecasting requires extrapolation from sequential observations. Language models, designed to discern intricate concepts within temporally correlated sequences, intuitively appear well-suited for this task. Hence, LLMs demonstrate proficiency in the domain of time series forecasting.

[0004] Currently, LLMs based on transformer model architectures use a mechanism called self-attention that enables learning of patterns and relationships over long sequences of data. LLMs have been predominantly used in language modelling using text tokens that are mapped to a vector space representation. This feature of LLMs can be leveraged for time series prediction which involves looking at patterns and trends that change over time. The data points from the time series domain are typically real numbers which can be quantized and scaled and then considered equivalent to the text tokens. This method of leveraging LLMs for time-series prediction involves the following steps:1. Pre-training: Pre-training an LLM on a massive dataset of text or code to learn longterm dependencies and patterns.PJN1051922. Fine-tuning: Fine-tuning the pre-trained LLM on a time series dataset to adapt the model's language understanding capabilities for time series prediction.3. Prediction: Using the fine-tuned LLM to predict future values of a time series based on historical data.

[0005] However, their application is limited to certain kinds of time series data. One such area that remains relatively unexplored is the prediction of network key performance indicators such as Downlink or Uplink Network Throughput, Downlink or Uplink Network Latency, Network Bearer Resource Utilization, Network Accessibility, Network Retainability, or Physical Resource Block (PRB) utilization, or the number of network servers or containers or pods and their associated capacities required to support one or more network functions, distributed allocation of such resources to support network resource requirements, etc.

[0006] Generally, in a wireless communication network, the resource allocation for a network session / flow is done based on resource requirements that are specified at the time of the start of the network session / flow. The cumulative resource allocation across different network sessions / flows across users represents the overall resource allocation in the network. Here the network resource allocation can represent the energy utilized in the network, or the bandwidth or latency budget allocated in network resource bearer, or the physical resources on a wireless communication channel, or virtualized hardware / software resources allocated to support network functions, or any combination of such allocations. ALML-based predictive techniques can be utilized based on past network data to estimate future network resource utilization and to help with dynamic resource allocation in the network to optimize the utilization of network resources. This work further extends such body of work by utilizing large language models to predict network KPI data. A combination of coarse-grain LLMs and a fine grain LLMs-based approach is suggested which helps in predicting network resource utilization for larger time steps and shorter time steps, respectively. The network resource utilization can be further refined based on an explicit known change in the network context (such as the start of a sports event). The network resource utilization prediction is then compared with live network KPI data to detect an anomaly in the prediction, model the anomaly and insert the anomaly to refine the network utilization prediction to produce the final network digital twin output data.

[0007] Such predicted models can then be utilized to proactively optimize the usage of network resources such as to minimize energy utilization using LLMs. Thus, this work proposes the utilization of LLMs to predict network KPI data, and to also optimize the future utilization ofPJN105192network resources. In particular, an example of network KPI modelling for Physical Resource Block (PRB) utilization is considered in this work. The overall network can comprise a distributed deployed of a combination of base-stations such as 4G eNodeBs, 5G gNodeBs, Open RAN-based disaggregated radio access network components (such as Central Units (CUs), Distributed Units (DUs) and Radio Units (RUs), future 6G-based access nodes, or access nodes based on alternative network access technologies. The predicted network KPI data it utilized to determine which of such access nodes and related network resources can be turned on or off or reconfigured or kept in an idle or sleep state to minimize energy utilization in the network. A wireless network can comprise a set of Coverage and Capacity cells operating at one or more frequency bands with different coverage range (radial and / or angular distance) using different wireless communication technologies. Coverage cells are typically always on to provide coverage in a given region. Capacity cells can be opportunistically utilized when there is increased network demand to augment the available capacity in a coverage cell, and they can be turned on or off or reconfigured dynamically. Coverage cells can also be turned on or off or reconfigured dynamically. For example, one coverage cell can be turned off if the range offered an alternative coverage cell can be reconfigured to transmit at a higher permitted power to provide adequate range corresponding to both coverage cells. This can be done when the network utilization becomes low enough that only one coverage cell is required in the system. Alternatively, a coverage or capacity cell can adapt the angular range of coverage by varying the direction and angular spread of the transmission beam, where the angular range can vary from a small directional angle to a large directional angle or to an omni-directional angle. The operation of coverage and capacity cells can consume a lot of energy in the network, so that dynamically turning such cells on or off or reconfiguring such cells can help in saving energy utilization in the system. Based on the configuration of a model for the network that represents different cells in the network, with the expected load based on the number of users in the geographical region as a function of time, a time-series prediction for network KPIs for each region covered by the cells is learned and predicted such that it is representative of dynamic network load in each of the cells. Based on the predicted network KPIs, the network can be reconfigured and optimized, such as to determine which resources (such as coverage or capacity cells), and to what degree to minimize energy utilization in the network.SUMMARY OF THE INVENTIONPJN105192

[0008] According to an embodiment of the disclosure, a method for real time prediction of a network Key Performance Indicators (KPIs) in an Al-enhanced communication network is disclosed. The method comprises the steps of receiving, by at least one Large Language Model (LLM) block, an input context information representing data for distinct time scales in form of a time-series waveform; generating a time series output based on the input context information; processing the generated output to enable automatic network reconfiguration in the AI-enhanced communication network by dynamically reconfiguring resource allocation in the network and by proactively reconfiguring network elements dynamically; and based on processing the generated output, computing the energy utilization in the Al-enhanced communication network.

[0009] Optionally, the network Key Performance Indicators (KPIs) includes a physical resource block (PRB) utilization in a network, or accessibility, or retainability, or a downlink throughput or a uplink throughput or a network latency, a network bearer throughput utilization for the RAN / transport / core network domains, a resource utilization in a data centre for Core / RAN network functions or a applications / services function or a container pods / clusters, end-to-end latency including the RAN / Transport / Core network domains, end-to-end throughput across the RAN / Transport / Core network domains.

[0010] Optionally, the reconfiguration of network elements can comprise turning on or off capacity cells or coverage cells in a wireless network, changing the transmit power or angular coverage of a coverage or capacity cell, or turning on or off wireless antennas or radio frequency (RF) processing paths in hardware, or turning on or off different cores in a server, or turning on or off or keeping in lower energy idle or sleep state for processing elements such as a CPU or cores in a server, or turning on or off dynamic scaling of software resources such as virtual functions or virtual machines or containers or pods.

[0011] Optionally, the resource allocation represents the energy utilized in the network, or the bandwidth or latency budget allocated in network resource bearer, or the physical resources on a wireless communication channel or the physical computation, communication and storage resources associated with a computing server or a data center, or virtualized hardware / software resources allocated to support network functions, or a combination of such resource allocations.

[0012] Optionally, the Large Language Model (LLM) block is an Artificial Intelligence (Al) system or a Machine Learning (ML) system or a state machine controller.PJN105192

[0013] Optionally, the time scales data in the input context information is arranged hierarchically.

[0014] Optionally, wherein the time series input waveform corresponds to the network key performance indicator (KPI) for a input time context window length in a given geographical region, where the time-window can be in a range of seconds or minutes or hours or days.

[0015] Optionally, the reconfiguration of a coverage or capacity cell can include increasing or decreasing the transmitted power within a permitted range for power to increase or decrease the range of the coverage or capacity cell, or to vary angular coverage of a cell.

[0016] Optionally, the time series output waveform corresponds to the network key performance indicator (KPI) for a future time window length in a given geographical region, where the time-series output can be based on a time-scale that is equivalent to one or more of processed input time- scales or at an alternative time- scale that is derived based on the available time-series output at different time-scales.

[0017] Optionally, additional input is integrated into the input context information to account for any days of the week and prior known anomalies.

[0018] Optionally, a coarse-grain level context drives the LLM block to produce a time series prediction at a fine-grain level, wherein the coarse-grain level context is obtained using another Al system or the Machine Learning (ML) system or the state machine controller.

[0019] Optionally, arranging the time scales data hierarchically implies that the input context information being obtained by consolidating the physical resource block (PRB) data at different time intervals wherein duration of the time intervals range from minutes to hours.

[0020] Optionally, energy utilization in the ALenhanced communication network is minimized by turning on or turning off a capacity-enhancing cell in a cellular network to optimize energy utilization in the ALenhanced communication network.

[0021] Optionally, performing explicit network anomaly insertion into the output time-series data based on an external input context related to the prior knowledge of one or more anomalies in the network.

[0022] Optionally, the prior knowledge of an anomaly in the network can be based on a prior known network effect or a live detected network anomaly.PJN105192

[0023] Optionally, detecting a network anomaly based on live observed network data to dynamically generate a model of the observed anomaly, where the modelling can be based on an explicit mathematical function or a non-linear neural network model, and to enable a digital twin generation of the output time-series data that adapts to the observed network data.

[0024] Optionally, an alternate network KPI data prediction is used with optional network anomaly insertion for a modelled network to generate a representative observed network timeseries KPI data that can be used to test and optimize the system for network KPI generation with anomaly adaptation and network optimization.

[0025] Optionally, dynamically observe the network KPI data and utilize the network KPI data to enable automated closed-loop reinforcement learning in the system.

[0026] According to another embodiment of the disclosure, an Artificial Intelligence (Al) system is disclosed. The system comprises a memory for storing an input context information, a communication interface for interacting with external entities, a processor configured to receive the input context information representing data for distinct time scales in form of a time-series waveform, generate an output based on the input context information, process the generated output to enable automatic network reconfiguration in the Al-enhanced communication network by proactively turning capacity cells ON or OFF dynamically, and based on processing the generated output, compute the energy utilization in the Al-enhanced communication network.

[0027] This summary is provided to describe select concepts in a simplified form that are further described in the detailed description. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

[0028] These and other objectives and advantages of the present disclosure will become more apparent when reference is made to the following description.BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The appended drawings depict typical embodiments of this disclosure, serving as illustrative examples rather than restrictive elements defining its scope. It's essential to recognise that the disclosure may encompass other equally effective embodiments beyond those depicted in the drawings. This acknowledgement underscores the versatility and potentialPJN105192applicability of the disclosure across a broader spectrum of variations and implementations, ensuring its adaptability to diverse contexts and requirements.

[0030] Figure 1 illustrates an LLM input-output model in accordance with an embodiment of the present disclosure;

[0031] Figure 2 illustrates a detailed model for forecasting using LLMs with multi-scale input as the context in accordance with an embodiment of the present disclosure;

[0032] Figure 3 illustrates the usage of Attention circuitry within LLMs to learn and predict time-series data with multi-scale input as the context in accordance with an embodiment of the present disclosure;

[0033] Figure 4 illustrates a detailed model for digital time anomaly adaptation to refine the predicted network KPI data forecasting in accordance with an embodiment of the present disclosure;

[0034] Figure 5 illustrates an additional model which takes into account for the days and prior known anomalies in accordance with an embodiment of the present disclosure;

[0035] Figure 6 illustrates a schematic closed-loop network configuration for network optimization in accordance with an embodiment of the present disclosure;

[0036] Figure 7 illustrates a schematic closed-loop automated reinforcement learning to refine the network KPI predictions, network anomaly adaptation, and network optimization in the system using automated feedback of live observed network data after network configuration in accordance with an embodiment of the present disclosure;

[0037] Figure 8 illustrates a schematic closed-loop automated reinforcement learning with automatic generation of network KPIs based on a configured network model with optional dynamic anomaly insertion to enable the testing and optimization of the network KPI prediction system in accordance with an embodiment of the present disclosure;PJN105192

[0038] Figure 9 illustrates a method for real time prediction of a network Key Performance Indicators (KPIs) in an Al-enhanced communication network in accordance with an embodiment of the present disclosure; and

[0039] Figure 10 illustrates an overall architectural framework of an Al system in accordance with an embodiment of the present disclosure.

[0040] Further, skilled artisans will appreciate that elements in the drawings are illustrated for simplicity and may not have necessarily been drawn to scale. For example, the flow charts illustrate the method in terms of the most prominent steps involved to help improve understanding of aspects of the present disclosure. Furthermore, in terms of the construction of the apparatus, one or more components of the apparatus may have been represented in the drawings by conventional symbols, and the drawings may show only those specific details that are pertinent to understanding the embodiments of the present disclosure so as not to obscure the drawings with details that will be readily apparent to those of ordinary skill in the art having the benefit of the description herein.DETAILED DESCRIPTION

[0041] The following description should be read with reference to the drawings, in which like elements in different drawings are numbered in like fashion. The drawings, which are not necessarily to scale, depict examples that are not intended to limit the scope of the disclosure. Although examples are illustrated for the various elements, those skilled in the art will recognize that many of the examples provided have suitable alternatives that may be utilized.

[0042] As used in this specification and the appended claims, the singular forms “a”, “an”, and “the” include the plural referents unless the content clearly dictates otherwise. As used in this specification and the appended claims, the term “or” is generally employed in its sense including “and / or” unless the content clearly dictates otherwise.

[0043] It is noted that references in the specification to “an embodiment”, “some embodiments”, “other embodiments”, etc., indicate that the embodiment described may include a particular feature, structure, or characteristic, but every embodiment may not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is contemplated that thePJN105192feature, structure, or characteristic may be applied to other embodiments whether or not explicitly described unless clearly stated to the contrary.

[0044] For the purpose of promoting an understanding of the principles of the disclosure, reference will now be made to the embodiment illustrated in the figures 1 through 10 and specific language will be used to describe the same. It will nevertheless be understood that no limitation of the scope of the disclosure is thereby intended, such alterations and further modifications in the illustrated system, and such further applications of the principles of the disclosure as illustrated therein being contemplated as would normally occur to one skilled in the art to which the disclosure relates.

[0045] The present disclosure focuses on prediction of PRB utilization time-series using hierarchical contexts. The advantages of using hierarchical models are leveraged by using multi-scale inputs for the model. The hierarchical context information helps in the feature extraction at a comparatively coarser and finer levels which is essential for precise predictions. Like, in case of weather predictions, the coarser level contexts help in giving the contextual information about the climate and the finer context gives information about the particular day’s weather. With both information combined, the model is more aware of the situation and makes an appropriate prediction.

[0046] Figure 1 illustrates an LLM input-output model where an input time series sequence of length L (comprising past and current data) is used to generate a future output time series window of length W. The LLM based on transformer model architecture uses a mechanism called self-attention that enables learning of patterns and relationships over long sequences of data.

[0047] Figure 2 illustrates a detailed model for forecasting using LLMs with multi-scale input as the context. The model incorporates hierarchical information into a input, which addresses the limitations of using a fixed time window. For example, with a 5-day window (a finer context), the model may lack a broader view of trends, while expanding the window to 14 days (a coarser context) may cause it to overlook finer details. To balance this trade-off, a hierarchical context is used, aggregating information from different time scales (such as multiple day ranges). This allows the model to consider both short-term details and long-term trends, improving its predictive performance. The time-series output can be produced at different time- scales to provide visibilities at different time -ranges that can help with planning for future resource allocation at different time-scales. For example, for an output window W atPJN105192a given time-scale can comprise W values spread over a certain period of time T, whereas another set of W values at a different scale that is slower by a factor of 2 would comprise W values spread over a larger period of time 2T. Based on the availability of information at different time-scales, an output window W can be generated at any of the set of time-scales that are generated, or an alternate time-scale by post-processing the time-series output that is available at different time- scales. Based on the prediction, the required amount or set of network resources can be selected. In the specific case of network energy minimization, the system can choose to turn on or off capacity cells in the region based on live predicted network demand. This can relate to turning on or off or reconfiguring radios or antennas that are utilized for wireless communication and scaling up or down of software resources (such as the number of active cores in a server, or the number of active servers or the number of instances of virtual functions or virtual machines or containers or pods) that are required.

[0048] Figure 3 illustrates the usage of attention circuitry within LLMs to capture time series behavior. Such attention circuitry can learn different forms of variations of time series across different attention circuits. Subsequently, for a given input context time series, an aggregate processing across the attention circuitry is used to generate future time series data.

[0049] Figure 4 illustrates a detailed model for digital time anomaly processing to refine the predicted network KPI data forecasting. The generated predicted network KPI data can be further refined optionally based on the knowledge of an explicit upcoming anomaly if such an explicit anomaly is specified. Such an anomaly can relate, for example, to the increased anticipated network usage during a sports event. The system can insert the expected behavior for such an anomaly, where the behavior associated with such an anomaly can be learned and predicted based on past available network data associated with such an anomaly. An explicitly-inserted anomaly can also relate to a live detected anomaly that is modeled in real-time and the model associated with the anomaly is inserted into the prediction. Subsequently, the predicted network data is compared information arriving from the live network to detect any anomalies. The behavior associated with the detected anomaly is observed and modeled dynamically. The modeled behavior is then inserted into the current predicted network data sequence to produce a digital twin output for the system that reflects information from live incoming network data. The behavior for a detected anomaly can manifest itself in various forms, such as a normal random process N(p,o) with mean p and variance c2, or it could manifest itself as an increasing or decreasing function of time with a noisy variance around it, or alternative mathematical forms, or utilizing a neural network to create a non-linear functional representation of thePJN105192anomaly. Once learned, the observed model for the detected anomaly is inserted into the predicted model to refine the prediction, where the observed model for the detected anomaly can represent an explicit mathematical function or a non-linear neural network model. A detected anomaly can then be chosen to be explicitly inserted into the explicit anomaly insertion stage, so that the detection of the detected anomaly is no longer required for the prediction of future context windows.

[0050] Figure 5 illustrates an additional model which takes into account for the days and prior known anomalies. The model is built on LLMs which adapts to the current trend by attending more to the deviations from the usual trend and that takes additional inputs like the day of the week and the type of anomaly associated with that day (the ones that are already known, this is an optional input). This can also be achieved by integrating another deep learning model with the LLM model or modifying the LLM architecture that acts as deviation control and adapts to the days and anomalies as required. This can be done as a post-processing task or as a pre-processing task. Alternatively, we can also have a state machine that lets us select the day as required. This way, the model adapts based on the additional inputs.

[0051] Figure 6 depicts the overall workflow for AI / ML-driven closed-loop network optimisation driven by observed network data and digital twin modelling, followed by proactive network optimization and control the predicted network data based on digital twin modelling is utilized to determine the optimal number of resources required for operations.

[0052] Figure 7 illustrates a schematic closed-loop automated reinforcement learning to refine the network KPI predictions and network optimization in the system using automated feedback of live observed network data after network configuration. This allows dynamic controllability of the network, along with dynamic observability of the network as well based on observed network KPIs, which enables dynamic automated closed loop adaptation in the network. The automated observability of network KPIs in the network enables the system to be optimized where the control actions in the system can be studied to verify if an action actually helped in improving the network KPIs in a desired direction. Corrective action can be taken in the system design if adverse effects are seen due to a control action. This helps in enabling automated reinforcement learning in this system with the observability and the controllability, so that human feedback is not a requirement to enable reinforcement learning in the system.PJN105192

[0053] Figure 8 illustrates a schematic closed-loop automated reinforcement learning with automatic generation of network KPIs based on a configured network model with optional dynamic anomaly insertion to enable the testing and optimization of the network KPI data prediction system. This enables the network KPI data prediction system to be tested for different scenarios for dynamic anomaly adaptation in the absence of a real network. This also helps the network optimization algorithm to be optimized in the presence of network anomalies.

[0054] For the experimental efforts related to this work, large language models (LLMs) are utilized for time-series prediction to predict network KPIs (key performance indicators) such as physical resource block (PRB) utilization in Al-enhanced 6G Networks served by a combination of coverage cells and capacity cells. An optimization algorithm processes the predicted time series data to enable automatic network reconfiguration to minimize energy utilization in the network by proactively turning capacity cells on or off or associated network hardware or software resources on or off dynamically. As the network utilization increases over time and crosses a certain threshold KI, then an additional capacity cell can be turned on if available to serve the same region. When the network utilization decreases at some future point in time beyond a certain threshold K2, then the capacity cell can be turned off. It is desirable to choose different values of the thresholds KI and K2, with KI > K2, SO that ping-pong effects can be avoided. For example, KI can be chosen to be a value of 85% of the overall capacity of the coverage cell, whereas K2can be chosen more conservatively at say 80% of the overall capacity of the coverage cell. The invention explores different approaches to predict PRB utilization using attention circuitry in an LLM framework, that learns time series behavior at different time-scales to predict future time series data based on an input context time-series data

[0055] Multiple inferential tasks have been carried out using LLMs, which essentially means that we are predicting without training the model. LLMs takes a context sequence of a certain maximum length and gives the prediction of the trend for the next mentioned number of time steps after the given context.

[0056] Dataset details: The created dataset mimics the actual behavior of the waveform and different basis waveforms such as sinusoidal functions or triangular waveforms or trapezoidal waveforms or rectangular waveforms or increasing functions or decreasing functions or concave functions or convex functions and random variations around such functions were used as the basis for forming the waveforms. The mean and variance for sampling from gaussianPJN105192processes were chosen to represent random variations in the basis functions. A combination of such basis functions with random variations was chosen to create the typical expected variations in network KPI data as a function of time during a day or during different days of a week. This was done to ensure that the synthetic waveforms duplicate some of the trends that would be expected in such data that would be representative of typical behavior in such a network. Different types of waveforms were generated for different regions that would be expected in different neighborhoods such as a commercial neighborhood or a residential neighborhood.

[0057] The dataset is created for 96 intervals per day. So, an hour is divided into 4 intervals of 15 minutes each. The dataset also accounts for the weekends and anomalies with sudden spikes and lows. The addition of random noise in the waveform is also used to generate dynamic variation in the data.

[0058] Experiments and results: The first experiment used pre-trained LLMs directly as a predictor of PRB utilization for the very next time step of the given input context. The pretrained LLM-based modeling was compared with a refined LLM-based model with fine-tuning, where the fine-tuning was performed based on expected KPI data for the network models under consideration with a combination of coverage and capacity cells. Time-series modeling networks based on RNNs (Recurrent Neural Networks) or LSTMs (Long-Short Term Memories) were also trained and tested on the same dataset. It was observed that the timeseries modeling with LLMs coupled with fine turning provided the best performance in the system.

[0059] It was observed that when the context length is limited to a short time-scale window, then the network models are not able to capture the trends over longer period such from a weekend to the next (given the number of intervals per day is high). This reasoning was verified with another experiment that considered a lesser number of intervals per day so that the input context has information of a longer period, and the LLMs in this case seemed to perform relatively better. Hence, multiscale -based processing was utilized in the system to provide network understanding and observability at different time-scales.

[0060] Another experiment conducted involved a CNN (Convolutional Neural Network) along with the LLMs. This addition was for improvising on the LLMs output by changing its prediction outputs by a small delta factor that is predicted by the CNN model which takes thePJN105192day and the type of anomaly as its input. In this manner, the output of the LLMs’ outputs is adjusted by taking the day and the type of anomaly of that day into consideration.

[0061] Figure 9 illustrates a method (900) for real time prediction of a physical resource block (PRB) utilization in an Al-enhanced communication network served by a combination of coverage cells and capacity cells.

[0062] Step 901 relates to receiving an input context information which contains data related to distinct time scales in form of a time-series waveform by the LLM block. Step 902 relates to generating an output based on the input context information.

[0063] Further, the method comprises of processing the generated output to enable automatic network reconfiguration in the Al-enhanced communication network by proactively turning capacity cells ON or OFF dynamically to reduce the energy utilization, as illustrated in step 903. Step 904 relates to the processed output of step 903 such that the method further comprises computing the energy utilization in the Al-enhanced communication network to find the amount of energy saving in the network. The above-described steps of figure 9 helps the system in taking decision in allocating physical resource block (PRB) and other relevant features in a communication network.

[0064] The present disclosure introduces a methodology for proactive time-series prediction by leveraging Al-based LLMs. LLMs specialize in capturing the patterns and features of sequential data. The approach is versatile, and its application can be extended across various domains. Apart from the telecommunications industry, where it can optimize network performance, it can also be used in sectors like agriculture, to forecast the optimal timing for crop growth based on weather conditions. Additionally, it holds potential for weather prediction itself, stock market forecasting, and improving efficiency in cloud data centres.

[0065] The present disclosure focuses on prediction of PRB utilization time-series using hierarchical contexts. The advantages of using hierarchical models are leveraged by using multi-scale inputs for the model. The hierarchical context information helps in the feature extraction at a comparatively coarser and finer levels which is essential for precise predictions. Like, in case of weather predictions, the coarser level contexts help in giving the contextual information about the climate and the finer context gives information about the particular day’s weather. With both information combined, the model is more aware of the situation and makes an appropriate prediction.PJN105192

[0066] The disclosure also targets to reduce the energy consumption by turning off the high throughput network elements (called capacity cells) when not required, by predicting the amount of PRB utilization.

[0067] The present disclosure also leverages the capability of attention circuitry to capture time-series variations to learn and predict future network behaviour (compared to learning random processes for mean KPI values in different time-windows during the day - such as averaged KPI values over 15-minute time intervals with 96 time windows and properties (such as the mean and variance) of the random process associated with each time-window.

[0068] One of the objectives of the present disclosure is to provide real-time predictions of PRB utilization for varying times throughout the day, accounting for both residential and industrial areas. In essence, the method predicts the usage of Physical Resource Blocks (PRBs) in wireless networks with hierarchical context that is acquired by using multiple scales of the input context.

[0069] Another objective of the present disclosure is to design multiple frameworks for adaptive prediction of the PRB utilization, that takes variations from the usual trends and observed anomalies into account, making the entire process proactive.

[0070] Yet another advantage of the present disclosure is to reduce energy consumption by turning off the capacity cells based on their predicted requirement by the model.

[0071] Additional advantage of the present disclosure is to produce digital twins of the PRB utilization waveform, thereby, creating a simulated dataset that mimics the actual network waveform behaviour.

[0072] Figure 10 illustrates a schematic diagram of a system 1000 according to an embodiment of the disclosure. The Al system 1000 includes a processor 1001, a communication interface 1002, and a memory 1003. The processor 1001, the communication interface 1002, and the memory 1003 may be connected to each other via a bus 1004. The bus 1004 may be a peripheral component interconnect (peripheral component interconnect, PCI) bus, an extended industry standard architecture (extended industry standard architecture, EISA) bus, or the like. The bus 1004 may be classified into an address bus, a data bus, a control bus, and the like. For ease of representation, the bus 1004 is represented by using only one line in Figure 10, but it does not indicate that there is only one bus or one type of bus. The processor 1001 may be a central processing unit (central processing unit, CPU), a network processor (network processor, NP),PJN105192or a combination of a CPU and an NP. The processor may further include a hardware chip. The hardware chip may be an application- specific integrated circuit (application-specific integrated circuit, ASIC), a programmable logic device (programmable logic device, PLD), or a combination thereof. The PLD may be a complex programmable logic device (complex programmable logic device, CPLD), a field -programmable gate array (field -programmable gate array, FPGA), generic array logic (Generic Array Logic, GAL), or any combination thereof. The memory 1003 may be a volatile memory or a non-volatile memory or may include a volatile memory and a non-volatile memory. The non-volatile memory may be a read-only memory (read-only memory, ROM), a programmable read-only memory (programmable ROM, PROM), an erasable programmable read-only memory (erasable PROM, EPROM), an electrically erasable programmable read-only memory (electrically EPROM, EEPROM), or a flash memory. The volatile memory may be a random-access memory (random access memory, RAM), and is used as an external cache.

[0073] The connecting lines shown in the various figures contained herein are intended to represent exemplary functional relationships and / or physical couplings between the various elements. It should be noted that many alternative or additional functional relationships or physical connections may be present in an embodiment of the subject matter.

[0074] The subject matter may be described herein in terms of functional and / or logical block components, and with reference to symbolic representations of operations, processing tasks, and functions that may be performed by various computing components or products. It should be appreciated that the various block components shown in the figures may be realized by any number of hardware components configured to perform the specified functions. For example, an embodiment of a system or a component may employ various integrated circuit components, e.g., memory elements, digital signal processing elements, logic elements, look-up tables, or the like, which may carry out a variety of functions under the control of one or more microprocessors or other control products. Furthermore, embodiments of the subject matter described herein can be stored on, encoded on, or otherwise embodied by any suitable non-transitory computer-readable medium as computer-executable instructions or data stored thereon that, when executed (e.g., by a processing system), facilitate the processes described above.

[0075] The foregoing description refers to elements or nodes or features being “coupled” together. As used herein, unless expressly stated otherwise, “coupled” means that onePJN105192element / node / feature is directly or indirectly joined to (or directly or indirectly communicates with) another element / node / feature, and not necessarily mechanically. Thus, although the drawings may depict one exemplary arrangement of elements directly connected to one another, additional intervening elements, products, features, or components may be present in an embodiment of the depicted subject matter. In addition, certain terminology may also be used herein for the purpose of reference only, and thus are not intended to be limiting.

[0076] The foregoing detailed description is merely exemplary in nature and is not intended to limit the subject matter of the application and uses thereof. Furthermore, there is no intention to be bound by any theory presented in the preceding background, brief summary, or the detailed description.

[0077] While at least one exemplary embodiment has been presented in the foregoing detailed description, it should be appreciated that a vast number of variations exist. It should also be appreciated that the exemplary embodiment or exemplary embodiments are only examples, and are not intended to limit the scope, applicability, or configuration of the subject matter in any way. Rather, the foregoing detailed description will provide those skilled in the art with a convenient road map for implementing an exemplary embodiment of the subject matter. It should be understood that various changes may be made in the function and arrangement of elements described in an exemplary embodiment without departing from the scope of the subject matter as set forth in the appended claims. Accordingly, details of the exemplary embodiments or other limitations described above should not be read into the claims absent a clear intention to the contrary.

Claims

PJN105192WE CLAIM:

1. A method (900) for real time prediction of a network Key Performance Indicators (KPIs) in an Al-enhanced communication network served by a combination of coverage cells and capacity cells, comprising the steps of:receiving (901), by at least one Large Language Model (LLM) block, an input context information representing data for distinct time scales in form of a time-series waveform; generating (902) a time series output based on the input context information; processing (903) the generated output to enable automatic network reconfiguration in the Al-enhanced communication network by dynamically reconfiguring resource allocation in the network and by proactively reconfiguring network elements dynamically; andbased on processing the generated output (904), computing the energy utilization in the Al-enhanced communication network.

2. The method as claimed in claim 1, wherein the network Key Performance Indicators (KPIs) includes a physical resource block (PRB) utilization in a network, or accessibility, or retainability, or a downlink throughput or a uplink throughput or a network latency, a network bearer throughput utilization for the RAN / transport / core network domains, a resource utilization in a data centre for Core / RAN network functions or a applications / services function or a container pods / clusters, end-to-end latency including the RAN / Transport / Core network domains, end-to-end throughput across the RAN / Transport / Core network domains.

3. The method as claimed in claim 1, wherein the reconfiguration of network elements can comprise turning on or off capacity cells or coverage cells in a wireless network, changing the transmit power or angular coverage of a coverage or capacity cell, or turning on or off wireless antennas or radio frequency (RF) processing paths in hardware, or turning on or off different cores in a server, or turning on or off or keeping in lower energy idle or sleep state for processing elements such as a CPU or cores in a server, or turning on or off dynamic scaling of software resources such as virtual functions or virtual machines or containers or pods.

4. The method as claimed in claim 1, wherein the resource allocation represents the energy utilized in the network, or the bandwidth or latency budget allocated in network resource bearer, or the physical resources on a wireless communication channel or the physical computation, communication and storage resources associated with a computing server or aPJN105192data center, or virtualized hardware / software resources allocated to support network functions, or a combination of such resource allocations.

5. The method as claimed in claim 1, wherein the Large Language Model (LLM) block is an Artificial Intelligence (Al) system or a Machine Learning (ML) system or a state machine controller.

6. The method as claimed in claim 1, wherein the time scales data in the input context information is arranged hierarchically.

7. The method as claimed in claim 1, wherein the time series input waveform corresponds to the network key performance indicator (KPI) for a input time context window length in a given geographical region, where the time-window can be in a range of seconds or minutes or hours or days.

8. The method as claimed in claim 1, wherein the reconfiguration of a coverage or capacity cell can include increasing or decreasing the transmitted power within a permitted range for power to increase or decrease the range of the coverage or capacity cell, or to vary angular coverage of a cell.

9. The method as claimed in claim 1, wherein the time series output waveform corresponds to the network key performance indicator (KPI) for a future time window length in a given geographical region, where the time-series output can be based on a time-scale that is equivalent to one or more of processed input time-scales or at an alternative time-scale that is derived based on the available time-series output at different time-scales.

10. The method as claimed in claim 1, further comprising integrating additional input into the input context information to account for any days of the week and prior known anomalies.

11. The method as claimed in claim 1, wherein a coarse-grain level context drives the LLM block to produce a time series prediction at a fine-grain level, wherein the coarse-grain level context is obtained using another Al system or the Machine Learning (ML) system or the state machine controller.

12. The method as claimed in claim 6, wherein arranging the time scales data hierarchically implies that the input context information being obtained by consolidating the physical resource block (PRB) data at different time intervals wherein duration of the time intervals range from minutes to hours.PJN10519213. The method as claimed in claim 1, further comprising minimizing energy utilization in the Al-enhanced communication network being carried out by turning on or turning off a capacityenhancing cell in a cellular network to optimize energy utilization in the Al-enhanced communication network.

14. The method as claimed in claim 1, wherein performing explicit network anomaly insertion into the output time-series data based on an external input context related to the prior knowledge of one or more anomalies in the network.

15. The method as claimed in claim 14, wherein the prior knowledge of an anomaly in the network can be based on a prior known network effect or a live detected network anomaly.

16. The method as claimed in claim 1, wherein detecting a network anomaly based on live observed network data to dynamically generate a model of the observed anomaly, where the modelling can be based on an explicit mathematical function or a non-linear neural network model, and to enable a digital twin generation of the output time- series data that adapts to the observed network data.

17. The method as claimed in claim 1, wherein an alternate network KPI data prediction is used with optional network anomaly insertion for a modelled network to generate a representative observed network time-series KPI data that can be used to test and optimize the system for network KPI generation with anomaly adaptation and network optimization.

18. The method as claimed in claim 17, wherein dynamically observe the network KPI data and utilizing the network KPI data to enable automated closed-loop reinforcement learning in the system.

19. An Artificial Intelligence system (1000) comprising:a memory (1003) for storing an input context information,a communication interface (1002) for interacting with external entities,a processor (1001) configured to:receive the input context information representing data for distinct time scales in form of a time- series waveform;generate an output based on the input context information;PJN105192process the generated output to enable automatic network reconfiguration in the AI-enhanced communication network by proactively turning capacity cells ON or OFF dynamically; andbased on processing the generated output, compute the energy utilization in the AI- enhanced communication network.

20. The system as claimed in claim 19, wherein the system is configured to perform the method steps as claimed in claims 1-18.

Images