Introduction

The large-scale deployment of renewable energy sources (RES), which is mainly made up of wind and solar photovoltaic (PV) energy, is at the core of the global shift to sustainable energy systems. These sources are frequently described as spatially dispersed and distant, i.e. offshore wind farms and large desert solar parks [1]. The effective and dependable bulk power delivery of these remote locations to larger load centers proves to be a big technical challenge [2]. Conventional alternating current (AC) grids have a low ability to transmit over a long distance because of the problem of reactive power losses, voltage fluctuation and synchronization. The High Voltage Direct Current (VSC-HVDC) technology utilizing Voltage Source Converter has become one of the most important enablers of the contemporary power systems providing clear benefits of integrating the RES [3]. These benefits are independent control of both active and reactive power, black-start, and capability to be connected to weak or passive AC networks [4]. Such a technological base makes it possible to advance beyond point-to-point HVDC connections to more complicated and flexible Multi-Terminal DC (MTDC) grids. The vision of MTDC grids has been to be linked grids capable of sharing the power generated by various geographically separated renewable generation centers and to redistribute and allocate power efficiently to other areas and hence increasing grid resilience and market flexibility [5].

The operation of a renewable-integrated MTDC grid presents complex challenges that span both control and protection domains. From a control perspective, the intermittent and stochastic behavior of RES, combined with the inertia-less characteristic of VSC stations, complicates the management of power flow, DC voltage stability, and dynamic power sharing among terminals. Traditional droop control schemes are often insufficient to address these issues [6].

Simultaneously, the MTDC system faces stringent protection challenges. DC fault can occur with extremely high rates of current side (di/dt) due to the low inductance of DC cables. Fault interruption is further complicated by the absence of natural current zero-crossings. Consequently, fault detection, classification and selective isolation must be executed very rapidly to prevent widespread converter blocking and potential catastrophic failure of the MTDC network [7].

Conventional solutions are often inadequate in this demanding environment. The widely used VSC control employs a linear Proportional-Integral (PI) controller, which is typically optimized for a specific operating point. However, its performance may degrade when operating across the wide and nonlinear range associated with variable renewable energy sources [8].

Traditional DC protection schemes, such as differential protection, can be limited by communication delays or they may not respond rapidly or selectively enough to support the large-scale meshed MTDC grids. A significant research gap exists in the prevailing siloed approach towards such problems. Although the issues of advanced control and intelligent protection schemes are commonly discussed separately, there is a gap in terms of a single, unified framework [9]. This research gap is an opportunity to capitalize on the flexibility and the ability to detect patterns of Artificial Intelligence (AI) and design a single hierarchical system that will ensure optimal operation and coordinate emergency fault correction [10].

The main aim of the research is to develop and test a single hierarchical AI-based control and protection system to increase the stability, efficiency, and resilience of the renewable-integrated MTDC systems. In order to reach this goal, the following specific contributions are used:

1. It is suggested to use a new three-level hierarchy of control (Primary, Secondary, Tertiary) with AI agents incorporated on the Primary and Secondary tiers in place of traditional control loops and to improve them.
2. A Deep Reinforcement Learning (DRL)-based Primary Controller is trained to be able to control local VSCs. This agent is informative of an ideal policy to substitute conventional inner current and outer power/voltage PI control, which allow a higher quality of dynamic response and flexibility.
3. A Secondary Coordinator based on AI is going to be developed to do optimal power sharing and DC voltage regulation. This module makes use of a centralized AI optimizer or a multi-agent system to dynamically operate system-wide setpoints, droop coefficients.
4. A Hybrid Advanced Fault Protection Scheme is provided. In this scheme, AI methods are used synergistically, i.e. Convolutional Neural Networks (CNNs) or Long Short-Term Memory (LSTM) networks to perform ultra-fast fault detection and classification, and a model-based algorithm to locate faults precisely.
5. An entire co-simulation validation framework is developed to validate the successful interplay of the AI-enhanced control hierarchy and intelligent protection scheme in a holistic and thoroughly considered situation of variability of the RES and DC faults under a wide range of operational and disturbance conditions.

The remainder of this paper is organized as follows. Section 2 presents a critical review of the available literature on the control of the category of MTDC, the application of AI in power electronics, and DC fault protection. Section 3 describes the architecture of the proposed hierarchical AI-driven control and protection scheme in detail. Section 4 explains the methodology, including the test system configuration, AI model development, and co-simulation framework. In Section 5 presents the simulation results along with a comprehensive discussion of system performance. Finally, Section 6 presents the conclusions and outlines directions for future research.

A. Literature Review

1. Control of MTDC Grids

Multi-Terminal VSC-HVDC systems are traditionally hierarchically organized to be professionally controlled to guarantee effective and stable operation. This hierarchy is usually of three tiers [11]. At the Primary level, local droop control is used on each converter to offer decentralized power sharing and DC voltage stabilization just after a disturbance has been experienced. The Secondary level is then employed to remove steady-state DC voltage errors or power-sharing errors by the droop characteristic, usually by a centralized or distributed controller which broadcasts corrective setpoints [12]. Lastly, the Tertiary level carries out system wide optimization like economic dispatch to set the optimal system power flow setpoints to the complete MTDC grid as dictated by the market conditions and grid requirements. In order to improve the performance over the traditional linear controllers, some sophisticated methods such as Model Predictive Control (MPC) have been explored. MPC provides better management of system constraints and control of multi-variables but has commonly been hampered by excessive computational complexity particularly in large networks [13]. A little bit later, the use of Artificial Intelligence (AI) methods has been examined. Fuzzy logic controllers have been developed to cope with the non-linearities of converter systems and Artificial Neural Networks (ANNs) have been applied to identify systems and as non-linear compensators. Early research on the field of Reinforcement Learning (RL) has shown that it has the potential to achieve adaptive control in power electronic systems because it is model-free, and it can be used to optimize long-term performance [14].

2. AI and Machine Learning in Power Electronics and HVDC

In the realm of AI implementation, Deep Reinforcement Learning (DRL) has received much interest in controlling power converters. The main benefit of DRL is that it is able to learn optimal control policies by interacting with a simulated environment, and does not need to have a clear and precise system model [15]. It has found application especially in non-linear, difficult-to-characterize systems such as VSC-HVDC terminals in a wide-range and uncertain environment, such as that dictated by intermittent renewable generation [16]. The agent is trained to optimize a specific reward form which may be a combination of several control goals such as reference tracking, oscillation damping, and constraint satisfaction. Parallel to it, supervised machine learning, mostly ANNs, is a popular technique used in auxiliary tasks. These involve dynamic system identification which trains a neural network model to act like the converters, and offline optimization of traditional controller parameters, to minimize reliance on expert knowledge and time-consuming manual performance optimization [17].

3. Fault Protection in VSC-MTDC Systems

The high-speed dynamics of DC faults pose a challenge to the protection of the MTDC grids. Fault currents increase at a very large di/dt and do not necessarily have a natural zero-crossing, so special methods are required to use [18]. These techniques are by blocking the IGBTs of the converters, which is easy but causes a temporary stop in transfer of power, or by using fast mechanical AC Circuit Breakers (CBs) with synchronized coordinates. More selective solution is the development of dedicated DC Circuit Breakers (DCCBs) which is also expensive. To activate these breaking devices many protection algorithms are suggested. Traveling-wave-based techniques make use of high-frequency content of voltage or current signals to measure fault direction and fault location at high speed but may be susceptible to noise and high sampling rates may be necessary [19]. Existing derivative techniques also have high speeds but are likely to malfunction when switching large loads or other transients. Differentiated protection programs provide great selectivity, but are fundamentally reliant on high-speed effective communication between end of line terminals, as well as establishing a possible failure point [20]. To address such drawbacks, AI and Machine Learning (ML) protection schemes are currently actively studied. Support Vector Machines (SVMs) and Random Forests are the techniques that have been used to classify the types of faults based on the features gained out of measured signals. Still more recently, various deep learning methods have been exploited to automatically derive spatio-temporal detail in raw or pre-processed voltage and current data to detect and classify faults more quickly and reliably [21]. Table 1introduces summary of current methods, their applications, and limitations.

4. Identified Research Niche

An overview of the literature analysis has shown that although considerable advances in the areas of advanced MTDC control and intelligent protection have been made individually, there still is a research gap. In the large majority of works, these two aspects, which are fundamental to each other, optimal system operation and emergency fault management, are considered independent design issues. Control schemes are commonly tested in the conditions of weak disturbances and protection algorithms are commonly tested without the dynamics of the control response. As a result, a clear absence of an integrated, holistic framework occurs. A combination of learning-based hierarchical control system, which allows real-time optimization and adaptation, with a coordinated AI-enhanced protection scheme, which is tolerant to rapid and reliable fault management, into one unified architecture is suggested as an innovative and required development. The research will fill this gap by presenting a single scheme in which AI agents control regular and abnormal operation as well as proactively support the protection system to provide a smooth transition and recovery during fault occurrences.

B. Proposed Hierarchical AI-Driven Control and Protection Scheme

1. Overall System Architecture

The general scheme of the proposed scheme is depicted in Figure 1. It is made up of a four-terminal VSC-MTDC grid that meshes a megawatt offshore wind power plant that is needed through Terminal 1 and a utility-scale photovoltaic (PV) farm that is needed through Terminal 3. Power sinks Terminals 2 and 4 are connected to strong mainland AC grids, which act as voltage supports. The hierarchical layers of control and protection that superimposes on the physical network layer, which includes cables, overhead lines, and VSC stations, controls the physical network layer. These layers work together to achieve power flow, DC voltage stability and ultra-fast and selective fault clearance.

2. Hierarchical Control Structure

Decoupling the control objectives at various timescales and system scopes is suggested using a hierarchical level of control structure.

3. Tertiary Level (Grid-Level Optimizer)

The Tertiary Level is a grid-level optimizer at a slow time scale. Periodically (i.e. every 15 minutes) a traditional or AI-enhanced Optimization Power Flow (OPF) algorithm is run according to market signals, forecasts of renewable generation and grid security requirements. The main objective is the economic dispatch, which is devised as a constrained optimization problem in order to minimize the overall generation cost or system losses. The level outputs are a series of optimal active power reference setpoints, , and DC voltage reference, , at each VSC terminal , and are transmitted to the Secondary Level.

Figure 1.

Figure 2.

Figure 3.

4. Secondary Level (Coordinator)

The Secondary Level is the main coordinator and it works on a timescale of seconds. It uses an AI-based controller (either in the form of a single deep neural network (DNN) or a multi-agent Deep Reinforcement Learning (DRL) system). It performs the dynamic reconciliation of the setpoints of the Tertiary Level to the conditions of the real-time systems. It is fed with internal measurements (e.g., ) and computes compensatory adjustments, and adaptive droop coefficients, . This guarantees optimal transmission of power between converters, no steady state voltage errors due to primary droop, and stability worldwide in the face of major disturbances. The control law may be in the form of a mapping that is learned by the AI as (1):

Figure 4.

Figure 5.

Figure 6.

Figure 7.

5. Primary Level (Local Controller)

The fast, local control of every VSC is done by the Primary Level, and acts on the millisecond's timescale. Traditional cascaded PI control loops are substituted by a special DRL agent, namely, Deep Deterministic Policy Gradient (DDPG) agent of each VSC. The state space of the agent, , consists of tracking errors and derivatives of active/reactive power and DC voltage as well as local AC-side current and voltage measurements. The converter direct modulation signal is the action, . The agent is provided with the reward function, , which is a penalty against tracking error and control effort as (2):

Figure 8.

Figure 9.

Figure 10.

Figure 11.

This model-free control has provided the capability of excellent dynamic performance, a natural ability to adapt to non-linearities and a strong reference tracking of the setpoints that the Secondary Level gives.

C. Advanced AI-Enhanced Fault Protection Scheme

1. Protection Architecture

The hybrid protection architecture is taken, which is a combination of the speed of local measurement-based decision-making and the selectivity offered by optional communication to verify. Every Intelligent Electronic Device (IED) on a line terminal function majorly on locally measured current and voltage derivatives.

Figure 12.

Figure 13.

2. Fault Detection & Classification Module

In this module, a lightweight Convolutional Neural Network (CNN) is used to transform a fixed length time-series window of the local current derivative signal . The CNN is offline trained to a large dataset with thousands of EMT simulation case scenarios due to different kinds of faults (Pole-to-Pole, Pole-to-Ground), locations, resistances, and inception angles. The network is a real time inference that gives an output in the form of a classification which is No Fault, pole-to- pole fault, or pole-to- ground fault.

Figure 14.

3. Fault Location Module

After detecting and classifying, a hybrid location module is launched. It uses AI classification result with a traveling-wave/impedance calculation algorithm based on a model. With the time difference between the arrival of the waves to two of the ends of the line, , the distance y of the fault can be estimated using the following equation; , where is the length of the line and is the wave propagation speed. The AI module will help to choose the right algorithm and check the measurement window.

Figure 15.

Figure 16.

Figure 17.

Figure 18.

4. Coordination with Control System

There is a critical firewall/coordination logic between the protection/control systems. When violation of fault is detected by the CNN with definitive fault, a tripping signal is transmitted to the corresponding DC Circuit Breakers (DCCBs) to isolate the minimal faulted area possible. At the same time, a safe and time-stamped mode-switching signal is reported to the Secondary AI Coordinator. The coordinator is notified of the new network topology (e.g. Line 1-2 out of service) by this signal. The Secondary AI is rapid at re-optimizing power flow and droop controls to the remaining healthy network, to achieve, and maintain, stable operation following a fault and also to reduce the effects of the contingency.

Method

A. Methodology and Simulation Setup

1. Test System Configuration

It is an engineered four-terminal, meshed MTDC grid with a symmetrical monopole set up that is set to a nominal DC link voltage . Offshore converter stations and onshore stations are two and connected with each other by overhead lines and submarine cables, with a length of 200 to 400 km. The parameters of the system such as cable per unit length, are set on the basis of the benchmark.

Figure 19.

Figure 20.

The renewable energy sources are incorporated as follows: Terminal 1 shall be linked to an elaborate aggregated model of an offshore wind power plant (WPP) with rated capacity of 900 MW. The WPP model includes a scaled-up converter-based wind turbine model, the mechanical power of which is provided by (3);

Figure 21.

where is the air density, A is the area covered, is the power coefficient, is the tip-speed ratio, is the pitch angle and vw is the wind velocity. The model is of a 33 kV collection grid consisting of a number of feeders. Terminal 3 is linked to a huge photovoltaic (PV) plant, which has a capacity of 600 MW. Single-diode equation is used to model the PV array as (4):

Figure 22.

Figure 23.

Figure 24.

Figure 25.

Figure 26.

where is the photocurrent, is the saturation current of the diode, is the ideality factor and and and are series and shunt resistances. The two renewable plants connect to the DC grid through special VSC stations and their switching dynamics are modeled by averaged models in training the controller and by detailed IGBT-based models in ultimate validation of EMT.

Figure 27.

Figure 28.

Figure 29.

Figure 30.

Figure 31.

2. AI Model Development and Training

Primary DRL Controller: Each of the VSC terminals has a Deep Deterministic Policy Gradient (DDPG) agent to substitute the traditional cascaded PI controllers. where is the state of the VSC at time as (5).

Figure 32.

Figure 33.

Figure 34.

Figure 35.

where , , and are the errors in the instantaneous active power, reactive power and DC voltage, respectively. is the integration of these errors and is the derivative. The action is the vector of modulation indices of the dq-axis voltages: .. The reward function rt i is aimed at reducing tracking error and managing effort as (6):

(6)

where are positive weightings coefficients. This agent is trained in a simulated environment within the OpenAI Gym interface with a stochastic variation of wind speed and solar irradiance based to and , with and being bounded random processes, to emulate real-world intermittency.

Secondary AI Coordinator: A centralized Artificial Neural Network (ANN) is implemented as the secondary coordinator. Its input vector is the global system state: , where and are vectors of measured powers and DC voltages from all terminals. The ANN outputs corrective terms and adaptive droop coefficients for each terminal. The ANN is trained offline via supervised learning, using data generated from a global optimizer that solves the constrained problem:

Secondary AI Coordinator: It is achieved by the means of a centralized ANN (artificial neural network) that is the second coordinator. The measured system state is the input state of the global system, Xsec as (7):

, (7)

are vectors of measured powers and DC voltages at each of the terminals. ANN is trained by an offline mode wherein the data which it produces is a result of a global optimizer which is addressing the constrained problem as (8):

subject to and . (8)

Protection AI Models: In the case of fault protection, a hybrid scheme is created that applies the AI-model. A 1D Convolutional Neural Network (CNN) is meant to detect and classify faults. In order to create its training set, thousands of EMT simulations are generated, injecting pole-to-pole (P2P) and pole-to-ground (P2G) faults at each 10 km along all DC lines. Fault resistance is varied and the inception angle is varied to . In both cases, a data window of raw current derivative signals di/dt at line-ends is obtained as (9):

. (9)

Where; is a set of labeled pairs, which consist of pairs of the form , and . Min-max scaling is employed to normalize all the data: . Another model-based algorithm applies the estimated type of fault identified by CNN and traveling wave principle, to the wave propagation velocity, to accurately determine the location.

Figure 36.

Figure 37.

Figure 38.

Figure 39.

Figure 40.

Figure 41.

Figure 42.

Figure 43.

Figure 44.

Figure 45.

Figure 46.

Figure 47.

Figure 48.

Figure 49.

Figure 50.

Figure 51.

Figure 52.

Figure 53.

Figure 54.

Figure 55.

Figure 56.

Figure 57.

Figure 58.

Figure 59.

Figure 60.

Figure 61.

Figure 62.

Figure 63.

Figure 64.

Figure 65.

Figure 66.

Figure 67.

Figure 68.

Figure 69.

Figure 70.

Figure 71.

Figure 72.

Figure 73.

Figure 74.

3. Integrated Co-Simulation Framework

An integrated co-simulation platform is created to be a real-time system that will verify the proposed scheme. The physical grid of the physical electromagnetic transient (EMT) model of the switched VSCs, cables, and renewable sources is simulated in a specialized power systems simulator (e.g., PSCAD/EMTDC). The control and protection algorithms that can be deployed through AI, namely the DDPG agents, the second ANN, and the protection CNN, are modeled and implemented in Python via TensorFlow/PyTorch. The two environments are aligned and connected through the standard of Functional Mock-up Interface (FMI). At a given time-step of the simulation the measurement signals are presented to the AI modules by the EMT model as inputs. The AI modules in turn feedback calculated control actions (modulation indices) and protection trip signals to the EMT model in order to actuate the VSCs and the circuit breakers. Such a closed loop arrangement forms a high fidelity, software in the loop testbed.

Figure 75.

Figure 76.

4. Case Studies and Performance Metrics

The integrated scheme is tested with respect to four different case studies with clearly defined quantitative measures.

Case 1: Normal functioning and renewable fluctuation. The example measures robustness when the variability in natural RES is considered. The performance parameters are the maximum DC voltage deviation , which measures the time needed to stabilize the active power to within a +-2 band after a change in the wind or solar input, the active power settling time which is the time required to settle the active power to a band following a step change in wind or solar input, and the damping ratio z of power oscillations approximated to the dominant modes of the system. Case 2: AC Side Faults. Three-phase fault and single-line-to-ground fault are used at the Point of Common Coupling (PCC) of an onshore VSC. The fault ride-through (FRT) feature is tested by ensuring that grid codes are met i.e. 150 ms. Recovery time is the time between fault clearance and the system fulfilling requirements: pu and pu. Case 3: DC Side Faults. This case study is a critical evaluation of the protection scheme. The most important metrics are: Fault Detection Time (minutes after the fault was generated till the CNN output has changed), Classification (fraction of correct classifications), Fault Location Error , , Total Fault Clearing Time Tclear (minutes since the fault was generated till the breaker has been opened), and the Post-Fault Recovery Time which is the amount of time it takes the grid to regain normal operation after the break. Case 4: N -1 Contingency (Converter Loss). Disconnection of one of the important VSC terminals (e.g., an onshore inverter) suddenly is modeled. System survivability is a binary metric that means that DC voltage is within a stable range ( pu). The efficiency of re-dispatch using power is measured as a percentage of the power loss that is actually recovered by the other terminals over a given period Tdispatch:

Figure 77.

Figure 78.

Figure 79.

Figure 80.

Figure 81.

Figure 82.

Figure 83.

Figure 84.

Figure 85.

Figure 86.

Figure 87.

Figure 88.

Figure 89.

Figure 90.

Figure 91.

Results and Discussion

The proposed hierarchical AI-based control and protection scheme are fully checked by an elaborate co-simulation system. It is a simulation environment, which takes the form of a four-terminal meshed MTDC grid with fluctuating renewable energy sources, based on the modified CIGRE B4 benchmark. Evaluation of system performance is done in four different operational conditions, namely, normal operation with renewable intermittency, AC side faults, DC pole-to-pole or pole-to-ground faults, and N-1 converter contingency conditions. The most important performance measures are the settling time, voltage variation, fault detection error, and recovery time, which are measured quantitatively. The findings in this section reveal that the integrated AI architecture is more efficient in terms of improving system stability, protection speed and general resilience in comparison with traditional control schemes.

The primary DRL controller convergence is shown, with the final average reward of 99.50 and convergence in episode 561, which is a measure of stable learning behavior. Protection AI module has very good fault classification features with a general accuracy of 98.93 with a high precision of P2P fault (98.32) and a recall of P2G fault of 98.80. The false positive rate of 0.631% is low which supports the validity of the suggested protection scheme in the differentiation of normal transient and fault conditions.

Figure 92. Figure 1. DRL Training Convergence and Fault Classification Performance.

Quantitative dynamism of the AI-based control is superior. The AI controller is 62.4 (0.320 s vs. 0.850 s) and 62.2 (3.1% vs. 8.2) faster in settling and voltage overshoot respectively than traditional PI control. The recovery time is increased by 56.9 percent during AC side faults (0.280 s instead of 0.650 s). These findings confirm the increased flexibility of the AI controller to renewable intermittency and disturbance rejection capacity.

Figure 93. Figure 2. Dynamic Performance Comparison Under Renewable Fluctuation and AC Faults.

The integrated protection scheme is ultra-fast in fault detection with an average of 0.61 ms and a 99th percentile of 1.48 ms, thus complying with the requirements of the MTDC protection. Both the speed and accuracy are shown by the total fault clearing time of 2.8 ms and the mean location error of 0.215 km. Figure 3(d) indicates that there is similarity in location accuracy at different fault distances with the largest error not exceeding 0.31 km at the distance of 200km.

Figure 94. Figure 3. DC Fault Protection Performance and Timing Analysis.

The hierarchical AI coordination will greatly increase resiliency of the system in converter loss. Voltage dip becomes smaller by 8.2% (0.92 pu compared to 0.85 pu), whereas recovery time decreases by 44.4% (0.279 s compared to 0.502 s). The loss of power in case of contingency lowers significantly by 26.9 % to 11.8 %, which indicates the usefulness of the secondary AI coordinator in the ideal redistribution of power and stabilization of voltages.

Figure 95. Figure 4. System Resilience During N-1 Converter Contingency.

The computational analysis determines that it is feasible to implement it in real-time. The 14.3 hours of total training are minimal compared to inference time (280 ms maximum protection CNN, 165 ms execution of the control loop). The overall memory footprint of 23.5 MB can be handled by the current hardware. These indicators justify the fact that the computational load does not affect the real-time functionality demands of the MTDC systems.

Figure 96. Figure 5. Computational Requirements and Memory Footprint Analysis.

Table 2 below gives a detailed quantitative overview of the main performance measures achieved during the simulation studies and which allow direct comparison of the conventional and proposed strategies.

The overall finding of the simulation results is that the suggested hierarchical AI architecture can lead to a high level of improvement in the performance in all of the metrics under assessment. The 62.4% change in settling time and 62.2% decrease in voltage overshoot proves the better dynamic response of the DRL based primary controller. The fact that the protection scheme has a very low latency of detection (0.61 ms mean) and a 98.93-percent accuracy is a great improvement to the traditional protection techniques, in which the latency of detection is generally 5-9 ms and the accuracy is 95-97 %. In addition, the 44.4% accelerated recovery in N-1 contingency and 56.1% drop in power loss confirm the efficiency of the secondary AI coordinator in terms of stability of the system in the case of severe disturbances. The computational analysis supports the practicability of the system, and the overall inference time is significantly less than the sub-milliseconds to support real-time operation of the systems in use. All these findings determine the suggested scheme as an overall solution to the stability, protection, and resilience of the renewable-integrated grids of the MTDC.

Conclusion and Future Work

The paper demonstrates the design and validation of a hierarchical AI-based scheme of the integrated control and protection of renewable-fed MTDC systems. The proposed architecture has been shown to provide improved dynamic performance, highly rapid and precise fault protection, and enhanced contingency resilience compared to traditional approaches. The seamless interaction between the AI-based control layers and the intelligent protection unit represents a significant step toward self-protective and adaptable MTDC grids. For practical implementation, the future work should focus on Hardware-in-the-Loop (HIL) validation using real-time digital simulators and power hardware prototypes to assess electromagnetic compatibility and systems latency. Additionally, the development explainable AI (XAI) methods is recommended to foster operational trust and facilitate compliance with grid codes. Investigations into cybersecurity measures for AI communication layers and the application of transfer learning to enable rapid adaptation to new network topologies are also identified as critical steps for deploying such intelligent systems in future power grids.