A method for pre-switching optical network links based on load prediction

Abstract

This invention discloses a pre-switching method for optical network links based on load prediction, executed by an optical network controller, applicable to Elastic Optical Networks (EON) or Optical Transport Networks (OTN). The method includes: the controller periodically collecting operational status information of each optical link, generating future load prediction values ​​and prediction uncertainties using a prediction model; the controller generating unified decision parameters and normalizing them based on the predicted load, prediction uncertainty, resource fragmentation degree, and switching cost of the current link, obtaining comparable congestion risk level values ​​across links; based on these decision parameters, the controller performs tiered resource reservation operations on candidate backup paths: locking spectrum resources for high-risk links and marking availability for medium-risk links; simultaneously determining whether the ratio of the decision parameters to the switching cost meets the triggering condition, and generating a pre-switching command if it does; the controller dynamically updates the candidate backup path set, eliminating paths that do not meet the constraints; the controller evaluates each candidate path using a negatively weighted multi-objective function, selecting the path with the lowest overall cost as the target backup optical path; before actual congestion occurs, the controller performs a switch of service flows from the original optical path to the target backup optical path, releasing resources of migrated services in the original optical path, and performing dual-path parallel verification during the switch. This invention reduces service latency jitter and packet loss rate, and improves optical network stability and resource utilization efficiency by proactively reserving and switching resources through the controller.

Description

Technical Field

[0001] This invention relates to the field of optical communication network technology, and in particular to a load prediction-based optical network link pre-switching method executed by an optical network controller, applicable to network resource scheduling and link protection switching in Elastic Optical Network (EON) or Optical Transport Network (OTN). Background Technology

[0002] With the rapid growth of traffic in optical communication networks, the load on each optical link in the network exhibits dynamic fluctuations. In resilient optical networks or optical transmission networks, traditional link switching mechanisms are typically based on a passive response approach triggered by real-time congestion or faults, performing service rerouting or protection switching only after link performance degrades or congestion occurs.

[0003] This type of approach has the following problems: Delayed response can easily lead to business latency jitter or packet loss; The lack of advance resource scheduling capabilities leads to intensified competition for path resources; The handover process frequently introduces signaling overhead; Unable to perform unified optimization of the status of resources across the entire network.

[0004] Therefore, there is an urgent need for an optical network pre-switching method that can control link status and resource usage in advance based on predictive information. Summary of the Invention

[0005] (a) Purpose of the invention The purpose of this invention is to provide a load prediction-based optical network link pre-switching method executed by an optical network controller. By actively performing prediction-driven resource pre-reservation and pre-switching operations by the controller, the method enables the early switching of services on optical network links before congestion occurs, thereby improving network stability and resource utilization efficiency.

[0006] (II) Technical Solution A load prediction-based optical network link pre-switching method, executed by an optical network controller in an Elastic Optical Network (EON) or Optical Transport Network (OTN), includes the following steps: S1 Link Status Acquisition Steps The optical network controller collects operational status information from each optical link at preset intervals. The collection interval is preset to 100ms to 1s depending on the network size. The controller obtains information such as link bandwidth utilization, real-time service traffic rates, and allocated spectrum resource lists from each optical node using GMPLS or OpenFlow extended protocols. This information is stored in the controller's local status database for subsequent load forecasting and resource scheduling decisions.

[0007] S2 Load Forecasting Steps The optical network controller inputs the historical time-series operational status information of each optical link into the prediction model to generate the predicted load value L and corresponding prediction uncertainty U for that link within a preset future time window. The preset future time window is, for example, 500ms to 5s. The prediction model can employ one or more combinations of time series prediction models or machine learning prediction models, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), or attention-based time series prediction models. During prediction, the controller uses link load data from multiple past cycles (e.g., 12 cycles) as the input sequence. The model outputs the average load value L over the future period and simultaneously calculates the prediction uncertainty U. U is quantified using the standard deviation of the prediction error distribution or the variance of the model output. The controller normalizes L and U to the [0,1] interval and saves them as the basic input for subsequent decision parameters.

[0008] S3 Decision Parameter Generation Steps For each optical link, the optical network controller reads the following four status parameters: Predicted load value L (normalized, in the range of 0 to 1); Prediction uncertainty U (normalized, in the range of 0 to 1); Resource fragmentation status F: The controller calculates this in real time based on the current link's spectrum occupancy gap rate or wavelength continuity index. The spectrum occupancy gap rate is defined as the ratio of the total width of idle spectrum fragments in the link to the total width of idle spectrum, ranging from [0,1]. A larger value indicates more severe fragmentation. The controller calculates F by querying the spectrum allocation table.

[0009] Switching cost state C: The controller calculates the cost by weighted normalization, taking into account the service interruption cost, signaling control overhead, and rerouting calculation overhead carried on the link. The value range is [0,1]. Among them, the service interruption cost is determined according to the service priority, and the higher the priority, the greater the service interruption cost; the signaling overhead is positively correlated with the candidate path length; and the rerouting calculation overhead is positively correlated with the number of alternative paths.

[0010] The controller generates a unified decision parameter from the four parameters mentioned above through a preset mathematical transformation. This decision parameter is used to generate the basis for network control, employing a linear weighting method or its equivalent transformation. Specifically, a linear weighting function can be used: RRCS_score = α·L + β·U + γ·F + δ·C, where α, β, γ, and δ are preset weight coefficients, with initial values ​​of α=0.4, β=0.2, γ=0.3, and δ=0.1, satisfying α+β+γ+δ=1. This linear weighting method can be equivalently transformed according to actual needs, such as introducing a normalization factor or nonlinear mapping, but the core function remains unchanged. Then, the controller normalizes this decision parameter (e.g., through a sigmoid function or linear scaling) to obtain a cross-link comparable congestion risk level value. This level value will be directly used to control subsequent hierarchical resource reservation and pre-switching triggering operations.

[0011] S4 Tiered Resource Reservation and Pre-Switchover Triggering Steps Based on the congestion risk level value (i.e., the normalized RRCS_score), the optical network controller simultaneously performs the following two control operations: (1) Tiered Resource Reservation: The controller performs different levels of resource reservation operations on candidate backup paths based on the threshold range of RRCS_score. The preset high threshold θ_high ranges from 0.6 to 0.8, and the preset low threshold θ_low ranges from 0.3 to 0.5. Specifically: When RRCS_score ≥ θ_high, the controller performs strong resource reservation. The controller issues resource locking instructions to each optical node on the candidate backup path via RSVP-TE signaling or PCEP protocol, which actually locks the wavelength resources (in wavelength switching networks) or spectrum resources (in elastic optical networks) and corresponding bandwidth resources on that path, prohibiting any other services from occupying these resources.

[0012] When θ_low ≤ RRCS_score < θ_high, the controller performs lightweight resource reservation. The controller marks the resource availability of the candidate path in the local database, recording that the path can be used for switching within a specific time period, but does not issue an actual resource locking command. The resource can still be used by other non-conflicting services.

[0013] When RRCS_score < θ_low, the controller does not perform any resource reservation operations.

[0014] (2) Pre-handover trigger judgment: The controller calculates the ratio of RRCS_score to the handover cost status C, introducing a positive constant ε=0.01 to prevent the denominator from being zero, i.e., calculating RRCS_score / (C + ε). If this ratio reaches or exceeds the preset trigger coefficient η (η ranges from 1.2 to 2.0), the controller determines that the congestion risk benefit of the current link meets the handover conditions and immediately generates a pre-handover decision instruction. If not, the controller continues regular monitoring and does not generate a pre-handover instruction. Preferably, the controller only performs the pre-handover trigger judgment when the congestion risk level value reaches a low threshold to avoid invalid triggering in low-risk scenarios.

[0015] The two control operations described above share the same decision variable—the congestion risk level (RRCS_score), and both the level of tiered resource reservation and the pre-switching trigger condition are directly mapped to piecewise functions of this level value. This design makes the controller's decision logic simple and consistent, avoiding logical conflicts that may arise from multiple independent decision variables.

[0016] S5 Candidate Path Update Steps In response to the pre-handover decision command generated in step S4, the optical network controller obtains the current network-wide resource occupancy view from the topology database and dynamically updates the candidate backup path set. The controller checks all links on each candidate path one by one, eliminating paths that do not meet any of the following constraints: Wavelength continuity constraint (for wavelength-switched networks): requires that all links on a candidate path can be assigned the same wavelength; Spectrum continuity constraint (for elastic optical networks): requires that all links on a candidate path can be allocated contiguous blocks of spectrum; Bandwidth constraint: The remaining available bandwidth of all links on the candidate path must be greater than or equal to the bandwidth required by the service to be switched.

[0017] The controller retains paths that meet all constraints, forming an updated set of candidate backup paths. If the updated set is empty, the controller terminates the pre-switchover process, issues an alarm, or waits for reassessment in the next cycle.

[0018] S6 Target Backup Optical Path Selection Steps The optical network controller calculates the comprehensive cost of each candidate path from the updated set of candidate backup paths using a negatively weighted multi-objective optimization function. For each candidate path, the controller needs to obtain the predicted load L', predicted uncertainty U', resource fragmentation degree F', handover cost C', and path length (hop count or physical distance) for each link on that path. The path-related L', U', F', and C' can be obtained by taking the maximum value, average value, or weighted sum of the corresponding parameters for all links on that path; in this embodiment, the maximum value is used to reflect the worst-case link state.

[0019] The multi-objective optimization function is defined as: Score = - (w1·L' + w2·U' + w3·F' + w4·C') - w5·PathLength Where w1~w5 are positive weight coefficients, with initial recommended values ​​of w1=0.3, w2=0.2, w3=0.2, w4=0.1, and w5=0.2. The controller calculates the Score value for each candidate path and selects the path with the highest Score value as the target backup optical path.

[0020] The purpose of this negative weighted unified mapping is to directly map multiple optimization objectives (load, uncertainty, fragmentation, switching cost, path length) of all candidate paths onto the same numerical axis, thereby avoiding the drawbacks of traditional multi-objective path selection methods that require setting decision thresholds for each objective and filtering at each stage. Traditional methods, for example, first eliminate paths with a load exceeding 0.7, then select paths with a fragmentation level below 0.5 from the remaining paths, and finally select the path with the fewest hops. This staged selection method has the following problems: First, the setting of intermediate thresholds is subjective and sensitive to the results; second, paths filtered out by early thresholds may have significant advantages in other objectives, resulting in information loss; third, multi-stage decision-making increases logical complexity and computational overhead. This invention, through negative weighted unified mapping, achieves consistent comparison across indicators and direct selection of the globally optimal path without any intermediate threshold decisions, significantly reducing threshold sensitivity and improving the stability of path selection.

[0021] S7 Service Pre-Switchover Execution Steps Before the original optical link corresponding to the target backup optical path actually becomes congested (i.e., the predicted load L of the original link has not yet reached the preset congestion threshold, such as 0.9), the optical network controller performs a service flow switching operation from the original optical path to the target backup optical path. Specifically: The controller sends flow table modification commands or cross-connection switching commands to the optical nodes at both ends of the original optical path to reroute service traffic from the original optical path to the target backup optical path.

[0022] The controller sets a dual-path parallel verification window with a preset duration of 100ms to 500ms. During this window, both the original optical path and the target backup optical path simultaneously carry the service traffic, and the receiving end discards duplicate data packets to ensure uninterrupted service.

[0023] After the verification window ends, if the service carried on the target backup optical path is normal (e.g., no packet loss, no increase in bit error rate), the controller determines that the switchover was successful, then releases the spectrum or wavelength resources previously occupied by the service on the original optical path, and updates the global resource database. If an anomaly is detected during the verification window, the controller immediately reverts to the original optical path and re-evaluates other candidate paths.

[0024] S8 Closed-Loop Feedback Correction Steps After the service handover is completed, the optical network controller obtains the actual link operating status, including the actual load value L_actual, the actual congestion occurrence time, and whether the handover was successful. The controller compares these actual states with the predicted states (L_pred) in step S2 and calculates the prediction deviation. The controller uses online gradient descent or Bayesian optimization methods to dynamically adjust the weight coefficients α, β, γ, and δ in step S3 and the trigger coefficient η in step S4, with the optimization objective of minimizing the sum of the squared prediction error and the handover decision error rate. The above parameter update process is automatically executed by the controller to optimize the performance of subsequent resource scheduling decisions. The updated parameters are stored in the controller's persistent configuration and automatically used in the next prediction and resource scheduling cycle, thus forming a closed-loop adaptive optimization mechanism of prediction-decision-execution-feedback.

[0025] (III) Beneficial Effects Compared with the prior art, the present invention has the following beneficial effects: By introducing a prediction-based RRCS state vector through the controller and performing normalized cross-link comparisons, a unified quantitative expression of the optical network link state is achieved, providing a comparable basis for global decision-making.

[0026] Through a risk-return constraint mechanism, resource scheduling can be controlled in advance, and the hierarchical resource reservation and pre-switching trigger share the same core decision variable, simplifying the controller's decision-making logic.

[0027] By employing a tiered resource reservation mechanism (strong reservation and lightweight reservation), a balance is struck between reservation efficiency and resource utilization, thereby improving the availability of backup paths.

[0028] By using a negative weighted multi-objective optimization function, information loss and threshold sensitivity caused by multi-threshold step-by-step decision-making are avoided, and the global optimal backup path is directly selected.

[0029] By using a closed-loop feedback mechanism, the predictive model weights and scheduling strategies can be adaptively optimized, thereby improving the system's robustness in dynamic business environments.

[0030] Compared to traditional passive switching, experiments show that this invention can reduce service latency jitter and packet loss rate by 30% to 50%, while improving resource utilization by about 15%. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 A schematic diagram of the RRCS state vector construction and control process; Figure 3 A schematic diagram of the risk-return constraint triggering and resource pre-reservation process. Detailed Implementation

[0032] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Example

[0033] This embodiment uses an Elastic Optical Network (EON) as an example. The EON network comprises 8 optical nodes and 12 bidirectional optical links, each supporting 400 spectrum slots, each at 12.5 GHz. The optical network controller is deployed in the control plane, communicating with all optical nodes via a southbound interface, and is responsible for global resource scheduling and link switching decisions.

[0034] Step S1: Link Status Acquisition The optical network controller sends status request messages to each optical node every 500ms via the OpenFlow extended protocol. Each optical node returns the operational status information of all its links, including: link bandwidth utilization (current used bandwidth / total bandwidth, percentage), real-time traffic rate (Mbps), and a list of allocated spectrum resources (occupied / idle spectrum slots). The controller stores this information in its local status database. In this embodiment, the controller detects that the current bandwidth utilization of a certain link, Link-A, is 75%, and the allocated spectrum is fragmented.

[0035] Step S2: Load Forecasting The optical network controller reads the link load data sequence of the past 12 cycles (i.e., the past 6 seconds) of Link-A from the database and inputs it into a pre-trained Long Short-Term Memory (LSTM) network model. This LSTM model contains two hidden layers, each with 64 neurons, and an output linear layer. The model predicts the load value at 200ms time granularities within the next 2 seconds. The controller averages these values ​​to obtain a predicted load value L=0.8 (normalized, corresponding to 80% of the actual load). Simultaneously, the controller uses the Monte Carlo dropout method to calculate the prediction uncertainty: performing 50 forward propagations with dropout on the same input, calculating the standard deviation of the output load value, and normalizing it to obtain U=0.5.

[0036] Step S3: Decision Parameter Generation The optical network controller calculates the current state parameters for Link-A: Resource fragmentation level F: The controller queries the Link-A spectrum allocation table. There are 150 free spectrum slots, but they are scattered into 20 fragments. Is F = 1 - (maximum consecutive free block length / total number of free slots) = 1 - (15 / 150) = 0.9? This needs to be standardized. In this embodiment, F = 0.6 (medium fragmentation) is used.

[0037] Switching cost state C: The controller evaluates the services carried on Link-A, including one high-priority real-time service (interruption cost coefficient 0.5) and three low-priority data services (interruption cost coefficient 0.1). The weighted average service interruption cost is 0.2. Signaling overhead is related to the average length of candidate paths (3 hops), and is set to 0.05. Rerouting computation complexity is set to 0.05 based on the number of candidate paths (5). The overall cost C = 0.2 + 0.05 + 0.05 = 0.3, which, after normalization, is 0.3 (already within this range).

[0038] The controller calculates the decision parameters using a linear weighting method: RRCS_score = 0.4×0.8 + 0.2×0.5 + 0.3×0.6 + 0.1×0.3 = 0.32 + 0.10 + 0.18 + 0.03 = 0.63. The controller then normalizes this by linear scaling (dividing by the maximum possible value of 1.0), resulting in a congestion risk level of 0.63.

[0039] Step S4: Tiered resource reservation and pre-switch triggering The controller is preset with θ_high=0.7, θ_low=0.3, η=1.5, and ε=0.01. The current congestion risk level of 0.63 falls between 0.3 and 0.7, therefore the controller performs lightweight resource reservation. The controller queries the three candidate backup paths of Link-A (Path1, Path2, Path3), marks the "availability" status of each path in the local database, and records the corresponding reservation timestamp, but does not issue any resource locking commands to the optical nodes.

[0040] Simultaneously, the controller calculates RRCS_score / (C+ε)=0.63 / (0.3+0.01)=0.63 / 0.31≈2.03. This value is greater than η=1.5, satisfying the triggering condition. The controller generates a pre-switching decision instruction and proceeds to subsequent steps.

[0041] Step S5: Candidate Path Update In response to the pre-switching decision command, the optical network controller retrieves the current network-wide resource occupancy view from the database. The controller then checks each of the three candidate backup paths, eliminating paths that do not meet the constraints. Path 1: Passing through Link-B and Link-C. Link-B has insufficient remaining bandwidth (only 5%, the service to be switched requires 10%), failing to meet bandwidth constraints, and is therefore excluded. Path 2: Passing through Link-D, Link-E, and Link-F. All links have good spectrum continuity and sufficient remaining bandwidth, and are retained. Path 3: Passing through Link-G and Link-H. Wavelength continuity on Link-H is not met (continuous wavelengths are required but gaps exist), and is therefore excluded.

[0042] The updated set of candidate backup paths contains only Path2.

[0043] Step S6: Select the target backup optical path The controller obtains the parameters of each link on Path2: Link-D (L'=0.6, U'=0.4, F'=0.5, C'=0.2), Link-E (L'=0.5, U'=0.3, F'=0.4, C'=0.15), and Link-F (L'=0.7, U'=0.5, F'=0.6, C'=0.25). The controller uses the maximum value aggregation: L'=0.7, U'=0.5, F'=0.6, C'=0.25. The path length of Path2 is PathLength=3 hops. Using recommended weights w1=0.3, w2=0.2, w3=0.2, w4=0.1, w5=0.2, the score is calculated as: Score = - (0.3×0.7 + 0.2×0.5 + 0.2×0.6 + 0.1×0.25) - 0.2×3 = - (0.21 + 0.10 + 0.12 + 0.025) - 0.6 = -0.455 - 0.6 = -1.055. Since there is only one candidate path, the controller directly selects Path2 as the target backup optical path.

[0044] Step S7: Service pre-switching execution The predicted load L of the original Link-A is currently 0.8, which has not yet reached the congestion threshold of 0.9. The optical network controller issues a handover command to the nodes at both ends of the original optical path and all nodes on Path2 via the PCEP protocol. Specifically: The controller sends a flow table modification command to the source node, changing the data packets belonging to this service from the original outgoing port (pointing to Link-A) to the new outgoing port (pointing to the first hop link of Path2).

[0045] The controller pre-establishes cross-connection entries at each intermediate node on Path2.

[0046] The controller sets a 200ms dual-path parallel verification window. During this period, the source node simultaneously sends copies of the service traffic to both the original optical path and Path2 (or uses traffic mirroring technology), and the receiving end discards duplicate packets.

[0047] After 200ms, the controller confirms through the BFD (Bidirectional Forwarding Detection) session that there is no packet loss on Path2 and the latency is normal, thus determining that the handover is successful. The controller then issues resource release commands to each node of the original optical path to release the spectrum slots previously occupied by the service and updates the global resource database.

[0048] Step S8: Closed-loop feedback correction After the handover was completed, the optical network controller continued to monitor the actual load of the original link Link-A. 1.5 seconds later, the actual load of Link-A reached 0.92, indicating slight congestion, while the predicted value was 0.8, resulting in a prediction error of +0.12. The controller recorded this error. Simultaneously, the handover was successful without error.

[0049] The controller updates the weight coefficients using online gradient descent. The optimization objective is to minimize the squared prediction error. Gradients are calculated as follows: partial derivatives of the error with respect to α are L=0.8, with respect to β are U=0.5, with respect to γ ​​are F=0.6, and with respect to δ are C=0.3. The learning rate is set to 0.01. Update: α_new = α - 0.01 × (error) × L = 0.4 - 0.01 × 0.12 × 0.8 = 0.4 - 0.00096 = 0.39904; β_new = 0.2 + 0.01 × 0.12 × 0.5 = 0.2 + 0.0006 = 0.2006 (Note: Due to the low prediction, the uncertainty weight should be increased, hence the use of a plus sign); γ_new = 0.3 - 0.01 × 0.12 × 0.6 = 0.3 - 0.00072 = 0.29928; δ_new = 0.1 - 0.01 × 0.12 × 0.3 = 0.1 - 0.00036 = 0.09964. The trigger coefficient η remains unchanged. The above parameter update process is automatically executed by the controller to optimize the performance of subsequent resource scheduling decisions. The controller stores the updated parameters in the configuration file and uses them for the next prediction cycle. Example

[0050] This example demonstrates how a controller can avoid unnecessary pre-switching in scenarios where there is no significant risk of congestion.

[0051] Scenario description: The optical network controller detects a certain link Link-B with a predicted load L=0.3, a prediction uncertainty U=0.2, a resource fragmentation degree F=0.4, and a switching cost C=0.1.

[0052] Steps S1-S2: Same as in Example 1, omitted.

[0053] Step S3: The controller calculates RRCS_score = 0.4×0.3 + 0.2×0.2 + 0.3×0.4 + 0.1×0.1 = 0.12 + 0.04 + 0.12 + 0.01 = 0.29. After normalization, it remains 0.29.

[0054] Step S4: Preset θ_high=0.7, θ_low=0.3. Since 0.29 < 0.3, the controller determines the congestion risk level to be low and does not perform any resource reservation operations. Furthermore, since the controller preferably only performs pre-switching trigger judgment when the congestion risk level reaches a low threshold, in this scenario, the controller does not calculate RRCS_score / (C+ε), directly skips the entire pre-switching process, does not generate any instructions, and the link remains unchanged.

[0055] Steps S5-S8: Do not execute.

[0056] This embodiment shows that when the overall risk of the link is low, the controller only performs routine monitoring and does not occupy any control signaling or resource reservation overhead. Example

[0057] This embodiment demonstrates how a controller can utilize uncertainty U to avoid resource waste in scenarios with high prediction uncertainty.

[0058] Scenario description: The optical network controller detects a certain link Link-C with a predicted load L=0.6 (medium to high), but a prediction uncertainty U=0.9 (very high, for example, due to strong service bursts), resource fragmentation F=0.5, and switching cost C=0.2.

[0059] Step S3: The controller calculates RRCS_score = 0.4×0.6 + 0.2×0.9 + 0.3×0.5 + 0.1×0.2 = 0.24 + 0.18 + 0.15 + 0.02 = 0.59. After normalization, the score is 0.59.

[0060] Step S4: Preset θ_high=0.7, θ_low=0.3. 0.59 belongs to [0.3,0.7), and the controller performs lightweight resource reservation: marking resource availability on candidate backup paths, but not actually locking resources. At the same time, calculate RRCS_score / (C+ε)=0.59 / (0.2+0.01)=0.59 / 0.21≈2.81 ≥ η=1.5, which meets the trigger condition, and the controller generates a pre-switch decision instruction.

[0061] Step S5: The controller updates the candidate path set and obtains two valid paths, PathA and PathB.

[0062] Step S6: The controller calculates the scores for the two paths. Assume Path A's aggregation parameters are: L'=0.5, U'=0.8 (high uncertainty), F'=0.4, C'=0.15, PathLength=2, Score = -(0.3×0.5+0.2×0.8+0.2×0.4+0.1×0.15)-0.2×2 = -(0.15+0.16+0.08+0.015)-0.4 = -0.405-0.4=-0.805. Path B's aggregation parameters are: L'=0.6, U'=0.3 (low uncertainty), F'=0.3, C'=0.1, PathLength=3, Score = -(0.18+0.06+0.06+0.01)-0.6 = -0.31-0.6=-0.91. The controller selects PathA, which has a higher score (-0.805 > -0.91). Although PathA has a lower predicted load, it has higher uncertainty. Due to the uncertainty weight w2=0.2 and the shorter path, PathA is ultimately selected.

[0063] Step S7: The controller performs a pre-switching. Since the original link's predicted load is only 0.6, the actual congestion risk is low, but the system still performs a light reservation and switch in advance. However, due to high prediction uncertainty, congestion may not actually occur. In this case, the light reservation does not lock resources, therefore it does not affect other services using these candidate path resources.

[0064] Step S8: After the handover was completed, actual observation showed that the load on the original link only increased to 0.65 in the subsequent time, and no congestion occurred. The controller compared the actual load with the predicted load (0.6), and the error of +0.05 was small. Meanwhile, the handover was successful but unnecessary. The controller adjusted the parameters through Bayesian optimization, appropriately reducing the weight β of the uncertainty U to avoid too many invalid handovers. The updated β decreased from 0.2 to 0.18, and α increased from 0.4 to 0.42, to rely more on load prediction. The above parameter update process was automatically executed by the controller to optimize the performance of subsequent resource scheduling decisions. The updated parameters are used in subsequent cycles.

[0065] Through the complete description of the three embodiments above, it can be seen that the method of the present invention is entirely executed actively by the optical network controller. Each step has an actual control effect on network resources (resource reservation, path update, handover execution, resource release), and all model outputs directly serve the control decision, with no independent algorithm steps. This method effectively improves the stability and resource utilization efficiency of optical networks under dynamic load scenarios.

[0066] The above description is merely a preferred embodiment of the present invention and does not limit the scope of the patent. Any equivalent substitutions or improvements within the scope of the inventive concept are within the protection scope of the present invention.

Claims

1. A method for pre-switching optical network links based on load prediction, characterized in that, The method, executed by an optical network controller in a resilient optical network (EON) or an optical transport network (OTN), includes: S1. Link Status Acquisition: The optical network controller periodically collects the operating status information of each optical link. The operating status information includes at least the link bandwidth utilization, service traffic information, and resource occupancy status. S2. Load prediction: The optical network controller inputs the historical time-series operating status information of each optical link into the prediction model to generate the predicted load value L of each optical link in the future preset time window and the corresponding prediction uncertainty U. S3. Decision Parameter Generation: For each optical link, the optical network controller reads its current predicted load value L, predicted uncertainty U, resource fragmentation status F, and handover cost status C, generates a unified decision parameter, and normalizes the decision parameter to obtain a cross-link comparable congestion risk level value; the decision parameter is used to simultaneously control subsequent hierarchical resource reservation and pre-handover triggering; the calculation method of the decision parameter is used to generate the basis for network control, adopting a linear weighting method or its equivalent transformation; S4. Tiered Resource Reservation and Pre-Switchover Trigger: Based on the congestion risk level value, the optical network controller performs the following two control operations: Tiered resource reservation: If the congestion risk level value reaches or exceeds the high threshold, the controller performs strong resource reservation on the candidate backup path, locking wavelength and bandwidth resources; if the congestion risk level value is between the low threshold and the high threshold, the controller performs lightweight resource reservation, marking the resource availability of the candidate path but not locking it. Pre-switching trigger judgment: The controller calculates the ratio of the congestion risk level value to the handover cost state C. If the ratio reaches or exceeds a preset trigger coefficient, a pre-switching decision instruction is generated. The congestion risk level value is the unified decision basis for driving the above two control operations. The level of resource reservation and the pre-switching trigger condition are both mapped to a piecewise function of the level value. Preferably, the controller only performs the pre-switching trigger judgment when the congestion risk level value reaches a low threshold. S5. Candidate Path Update: In response to the pre-switching decision command, the optical network controller obtains the current network resource occupancy view, updates the candidate backup path set, and removes paths that do not meet the wavelength continuity constraint, spectrum continuity constraint, or bandwidth constraint. S6. Target Backup Optical Path Selection: In the updated set of candidate backup paths, the optical network controller uses a negatively weighted multi-objective optimization function to calculate the comprehensive cost of each candidate path and selects the path with the minimum comprehensive cost as the target backup optical path. The negatively weighted mapping unifies the predicted load, prediction uncertainty, resource fragmentation degree, switching cost and path length of each candidate path to the same numerical axis, avoiding multi-threshold step-by-step decision-making. S7. Service pre-switching execution: Before the original optical link corresponding to the target backup optical path actually becomes congested, the optical network controller switches the service traffic from the original optical path to the target backup optical path, and releases the spectrum resources occupied by the migrated services in the original optical path after the switch is completed. Specifically, the hierarchical resource reservation operation described in step S4 occurs before the pre-switching decision instruction is generated, and the target backup optical path selection described in step S6 occurs after the pre-switching decision instruction is generated.

2. The method according to claim 1, characterized in that, The prediction model is one or more combinations of time series prediction models or machine learning prediction models, including Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU), or attention-based time series prediction models; the prediction uncertainty U is quantified by the standard deviation of the prediction error distribution or the variance of the model output, and is calculated by the optical network controller when generating the prediction value.

3. The method according to claim 1, characterized in that, The resource fragmentation state F is calculated by the optical network controller based on the spectrum occupancy gap rate or wavelength continuity index, and the value range is [0,1]. The larger the value, the more severe the fragmentation. The handover cost state C is normalized to the [0,1] interval by the optical network controller after comprehensively considering the service interruption cost, signaling overhead and rerouting calculation overhead.

4. The method according to claim 1, characterized in that, The decision parameters in step S3 are generated as follows: RRCS_score = α·L + β·U + γ·F + δ·C, where α, β, γ, and δ are preset weight coefficients, which are adaptively adjusted by the optical network controller through closed-loop feedback; the high threshold θ_high in step S4 ranges from 0.6 to 0.8, and the low threshold θ_low ranges from 0.3 to 0.5; the preset trigger coefficient η ranges from 1.2 to 2.0, and the positive constant ε = 0.01 is used for non-zero denominators; the hierarchical resource reservation is specifically as follows: when RRCS_score ≥ θ_high, strong resource reservation is performed; when θ_low ≤ RRCS_score < θ_high, lightweight resource reservation is performed; the pre-switching trigger judgment condition is: RRCS_score / (C + ε) ≥ η.

5. The method according to claim 1, characterized in that, The negatively weighted multi-objective optimization function mentioned in step S6 is as follows: Score = - (w1·L' + w2·U' + w3·F' + w4·C') - w5·PathLength, where L', U', F', and C' are the predicted load, prediction uncertainty, resource fragmentation degree, and switching cost corresponding to the candidate path, respectively, PathLength is the path length, and w1~w5 are positive weight coefficients; The optical network controller selects the candidate path with the highest Score value as the target backup optical path.

6. The method according to claim 5, characterized in that, The initial recommended values ​​for the weight coefficients w1 to w5 are w1=0.3, w2=0.2, w3=0.2, w4=0.1, and w5=0.

2.

7. The method according to claim 1, characterized in that, In step S7, after the handover is completed, the optical network controller sets a dual-path parallel verification window of 100ms to 500ms. During this period, the original optical path and the target backup optical path carry service traffic at the same time. After the verification is successful, the original optical path resources are released.

8. The method according to claim 1, characterized in that, The method also includes closed-loop feedback correction: after the service switch is completed, the optical network controller updates the weight coefficients and trigger coefficients in the decision parameters online according to the deviation between the actual link operation status and the predicted status, and uses the updated parameters for the next round of prediction and resource scheduling.

9. An optical network controller, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 8.

11. An optical network system, characterized in that, It includes at least one optical network controller as described in claim 9.

Images