AI, Digital Engineering and Emerging Technologies

Introduction and Context

Electrical engineering projects increasingly combine physical infrastructure with software, data, communication networks and intelligent decision-making. A modern dissertation may involve a photovoltaic inverter, battery-energy-storage system, smart meter, electric vehicle charger, autonomous robot, industrial controller or power-system simulator. In each case, engineering quality depends not only on circuit and system theory but also on the quality of data, the validity of models, cybersecurity, explainability and responsible deployment.

Artificial intelligence (AI) is useful when a system must identify patterns, estimate unmeasured variables, classify operating conditions, predict future behaviour or optimise decisions. It is not automatically the best solution. A physically derived model, conventional controller or carefully designed experiment may be safer, cheaper and more interpretable. The appropriate engineering approach is therefore hybrid: use physical knowledge, data-driven methods and professional judgement together.

This chapter examines AI, machine learning, digital twins, automated design, cloud-connected engineering and emerging applications in smart grids, power electronics, the Internet of Things (IoT), autonomous systems and electrification.

Learning Objectives

After studying this chapter, a student should be able to:

1. Explain the roles of AI, machine learning, digital twins and digital engineering in electrical projects.
2. Select suitable data, models and performance measures for an engineering problem.
3. Distinguish training, validation and test data and recognise common sources of bias and leakage.
4. Evaluate explainability, uncertainty, safety, cybersecurity and regulatory concerns.
5. Describe applications in smart grids, power electronics, IoT, autonomous systems and electrification.
6. Design a defensible AI-enabled dissertation methodology and communicate its limitations.

1. Digital Engineering as a Systems Approach

Digital engineering is the coordinated use of digital models, simulation, data, software, communication and lifecycle information to design, operate and improve engineered systems. It differs from simply using a computer program. A digital-engineering process maintains traceability between requirements, models, components, tests, measurements and decisions.

A useful lifecycle is:

Requirement → model → simulation → prototype → measurement → verification → deployment → monitoring → redesign

The same physical system may be represented at several levels:

• Component level: semiconductor switch, sensor, motor winding or battery cell.
• Subsystem level: inverter, motor drive, protection system or battery-management system.
• System level: microgrid, electric vehicle, factory or transmission network.
• Enterprise level: assets, maintenance, energy markets, regulations and users.

Model fidelity must match the engineering question. A detailed electromagnetic-transient model may be essential for switching behaviour but unnecessary for annual energy planning. Conversely, a simplified steady-state model may be unsuitable for semiconductor-stress prediction.

Model verification and validation

Verification asks, “Was the model implemented correctly?” Validation asks, “Does the model represent the real system adequately for its intended use?” A model can be mathematically correct yet physically inappropriate. Validation should specify operating range, sensor accuracy, parameter uncertainty and acceptable error.

For measured values yᵢ and predictions ŷᵢ, common regression measures include:

Here n is the number of observations and yᵢ and ŷᵢ have the same units. MAE represents average absolute error, while RMSE penalises large errors more strongly. Both assume that the observations being compared are meaningfully paired. Percentage measures can be misleading when yᵢ is zero or close to zero.

For a binary classifier, with true positives TP, true negatives TN, false positives FP and false negatives FN:

Accuracy can be inappropriate for highly imbalanced data. In a protection or fault-detection application, recall may be prioritised because missed faults can be more harmful than false alarms.

2. Artificial Intelligence and Machine Learning

AI is the broad field of systems performing tasks associated with human intelligence, such as perception, reasoning, prediction and decision-making. Machine learning (ML) is a family of methods in which a model learns relationships from data rather than being explicitly programmed for every case.

Main learning paradigms

Supervised learning uses labelled examples. Regression predicts continuous quantities such as load, temperature or battery state of charge. Classification predicts categories such as normal, faulty or overloaded.

Unsupervised learning identifies structure without labels. Clustering can group similar load profiles or operating conditions. Dimensionality-reduction methods can help visualise high-dimensional sensor data, but reduced coordinates do not automatically have physical meaning.

Reinforcement learning learns actions through interaction with an environment and a reward signal. It may be useful for energy management, but exploration in a live power system can be unsafe. Simulation, constrained action spaces and supervisory protection are therefore important.

A simple supervised-learning model

A linear regression model can be written as:

where ŷ is the predicted output, xⱼ are input features, βⱼ are fitted coefficients and β₀ is the intercept. The model assumes that a linear approximation is adequate over the intended operating range and that the data are sufficiently representative. Coefficients may be statistically unstable when features are strongly correlated.

Many models minimise a loss function. For squared-error regression:

The loss is an objective for fitting; it is not by itself evidence that the model will perform well on new data.

Training, validation and testing

The dataset should be divided before final evaluation. The training set fits model parameters. The validation set supports choices such as model type, feature selection and hyperparameters. The test set is reserved for an unbiased estimate of final performance. If time-series observations are randomly shuffled, future information may enter the training set and produce unrealistically good results. Chronological splitting or rolling-origin evaluation is usually more defensible.

Data leakage occurs when information unavailable at decision time enters the features or labels. Examples include using a post-fault signal to predict the fault, scaling all data before splitting, or allowing repeated measurements from the same physical event to appear in both training and test sets.

Data quality

Engineering data may contain missing values, sensor drift, quantisation, communication delays, duplicated records, outliers, inconsistent units and changing operating regimes. A responsible workflow includes:

1. Define the measurement context, sampling interval and units.
2. Check timestamps, synchronisation and sensor plausibility.
3. Identify missingness and determine whether it is random or systematic.
4. Detect outliers using physical limits and statistical methods.
5. Document filtering, interpolation and imputation.
6. Preserve raw data and maintain a versioned processing pipeline.
7. Test performance separately across relevant regimes, seasons, devices and users.

A high score on historical data may reflect memorisation, stable laboratory conditions or a hidden proxy for the target. Generalisation must be demonstrated using independent conditions.

3. Explainability, Uncertainty and Responsible Use

Explainability is the extent to which a human can understand why a model produced an output. A physically meaningful coefficient, decision tree or rule may be intrinsically interpretable. For complex models, post-hoc methods can estimate feature influence, but an explanation is not necessarily a causal explanation.

Engineering decisions require uncertainty, not only point predictions. A forecast of 10 kW without an uncertainty estimate may be inadequate for reserve scheduling. Prediction intervals, confidence intervals, ensembles or probabilistic models can communicate uncertainty, provided their coverage is validated.

Important limitations include:

• Distribution shift: operating conditions after deployment differ from training conditions.
• Concept drift: the relationship between variables changes over time, such as after equipment ageing.
• Class imbalance: rare but important faults are underrepresented.
• Correlation versus causation: a predictive feature may not be a controllable cause.
• Opacity: complex models may be difficult to audit.
• Overfitting: a model learns noise rather than transferable structure.
• Automation bias: operators may trust an apparently authoritative output.

A safety-critical system should retain independent protection and a safe fallback. AI should normally advise, estimate or optimise within constraints; it should not silently bypass certified protection functions.

4. Digital Twins

A digital twin is a digital representation of a physical asset or process that is connected to the asset through data and is used for monitoring, prediction, diagnosis or optimisation. A static CAD model is not necessarily a digital twin. A twin normally includes:

• a defined physical asset and purpose;
• geometric, electrical, thermal or behavioural models;
• real or near-real-time data connections;
• parameter updating or state estimation;
• lifecycle records and version control;
• validation criteria and a decision workflow.

For a battery, the twin may estimate state of charge (SOC), state of health (SOH), temperature and remaining useful life. SOC is not directly measured; it is inferred from current, voltage, temperature and a model. A simplified coulomb-counting relation is:

Here Qₙ is nominal charge capacity in ampere-hours, I is current under a declared sign convention, η is an efficiency factor, and time must be expressed consistently with the current unit. The formula is subject to sensor bias, uncertain initial SOC, temperature effects, capacity ageing and charging/discharging efficiency. Therefore, practical estimation often combines coulomb counting with voltage models or observers.

A twin should not be judged only by visual similarity. Its accuracy, latency, update frequency, uncertainty and decision value must be assessed. A highly detailed twin that cannot run fast enough for its intended application may be less useful than a simpler validated model.

5. Automated Design and Optimisation

Automated design searches a defined design space subject to constraints. Applications include conductor sizing, filter design, motor geometry, converter control parameters, antenna configuration and microgrid scheduling.

A generic constrained optimisation problem is:

The vector x contains design variables, f is the objective, gⱼ and hₖ are inequality and equality constraints, and xᴸ and xᵁ are lower and upper bounds. The objective and constraints must reflect engineering requirements such as efficiency, cost, thermal limits, voltage quality, reliability and manufacturability.

A design that minimises cost alone may violate safety or lifetime requirements. Multi-objective methods often produce a Pareto front: a set of designs for which improving one objective requires worsening another. The final selection remains an engineering decision.

Surrogate models can approximate an expensive simulation. However, they must be tested in the design region, especially near constraint boundaries. Optimisation should include sensitivity analysis: if a small parameter change causes a large performance change, manufacturing tolerances and uncertainty become central design concerns.

6. Cloud-Connected Engineering and IoT

An IoT-enabled electrical system typically contains sensors, signal conditioning, local computation, communication, cloud or edge services, analytics and an actuation or reporting layer. Edge computing processes data near the asset, reducing latency and dependence on network availability. Cloud systems support storage, fleet-level analytics and collaborative access.

A robust architecture separates functions:

• Sensing: measurement range, calibration, isolation and sampling.
• Edge control: deterministic and safety-critical functions.
• Communication: authenticated and encrypted data transfer.
• Cloud analytics: fleet comparison, model training and dashboards.
• Governance: access control, retention, audit and incident response.

Connectivity introduces attack surfaces. Threats include credential theft, spoofed measurements, denial of service, malware, insecure firmware and unauthorised remote commands. Good practice includes least-privilege access, secure boot, signed updates, network segmentation, key management, logging, vulnerability management and a manual or local fallback.

The sampling rate should relate to the phenomenon being studied. A slow environmental trend does not require the same rate as a power-electronic switching transient. Sampling, communication and storage choices should be justified by bandwidth, latency, aliasing, computational load and control requirements.

7. Applications in Emerging Electrical Technologies

Smart grids

Smart grids integrate distributed generation, storage, flexible demand, advanced metering, communications and automation. AI applications include load forecasting, renewable-power forecasting, non-technical-loss detection, outage localisation, voltage optimisation and demand response.

Forecast errors affect reserve and dispatch decisions. A model must therefore be evaluated across weather, seasons, holidays and unusual events. Forecasting should be integrated with physical network constraints: an accurate load prediction is not sufficient if the resulting schedule causes thermal overload or unacceptable voltage.

Power electronics

Power converters operate at fast time scales and are sensitive to switching, parasitics, thermal conditions and electromagnetic interference. ML can support fault diagnosis, parameter estimation, remaining-life prediction and modulation optimisation. It should complement, not replace, well-established control and protection principles.

Training data must include component ageing, load changes, temperature variation and sensor faults. A classifier trained only on healthy laboratory waveforms may fail in field conditions. Hardware-in-the-loop testing is valuable because it exposes algorithms to realistic timing and interfaces without immediately risking a full-power system.

Autonomous systems

Autonomous vehicles, drones and mobile robots combine perception, localisation, planning and control. Their engineering assurance requires defined operating domains, sensor redundancy, fail-safe states and verification of interactions between learned perception and deterministic control. A navigation model that works in clear daylight may not be reliable in rain, dust or sensor obstruction.

Electrification

Electrification of transport, heating and industry increases demand for power converters, batteries, charging infrastructure and distribution-network reinforcement. AI can coordinate charging, predict battery degradation and improve maintenance. It can also create inequitable outcomes if tariffs or charging access are optimised without considering vulnerable users and grid constraints.

8. Illustrative Case: Predictive Maintenance for a Solar Inverter

Background. A distribution-connected solar farm experiences intermittent inverter trips. Operators possess voltage, current, temperature, ambient-condition and alarm records, but labels identifying the true cause are incomplete.

Problem. The project team proposes an ML classifier to predict whether an inverter will trip within the next hour. The objective is early maintenance notification, not autonomous protection.

Analysis. The team first defines the prediction time and excludes signals recorded after a trip. It checks timestamps, missing data, firmware changes and inverter identity. Because trips are rare, accuracy is rejected as the sole metric. Recall, precision, false-alarm rate and performance by inverter and weather condition are reported.

Approach. A transparent baseline using logistic regression is compared with a tree-based model. Data are divided chronologically: earlier months for training, a subsequent period for validation and the final period for testing. Class weighting is considered, but threshold selection is based on maintenance capacity and the relative cost of missed trips and false alarms. Features include rolling temperature statistics, recent alarm counts and power-quality indicators available before the forecast time.

The system displays the predicted risk, influential features and an uncertainty or confidence indicator. It does not issue a trip command. A conventional protection system remains independent. Alerts are first evaluated in shadow mode, where predictions are logged without operational action.

Outcome. Suppose the final test contains 2,000 hourly windows, of which 80 precede a trip. The model identifies 64 of these and produces 96 false alarms. Then:

The results indicate that 80% of impending trips were detected, but only 40% of alerts corresponded to a trip. Whether this is acceptable depends on the cost of inspections, the cost of missed failures and the available maintenance workforce. The project should not claim general validity beyond the tested inverter types, seasons and operating conditions.

Learning points. The strongest contribution is not selecting the most complex algorithm. It is defining the prediction task, preventing leakage, preserving an independent test period, reporting operationally meaningful metrics and designing a safe deployment process.

9. Dissertation Methodology and Professional Practice

An AI-enabled electrical dissertation should contain a clear chain of evidence:

1. Research question: state the engineering decision or uncertainty being addressed.
2. System boundary: identify equipment, users, interfaces and excluded effects.
3. Baseline: establish a physical, statistical or conventional-control reference.
4. Data plan: define sensors, sampling, labels, units, access rights and quality checks.
5. Model plan: justify the algorithm and its assumptions.
6. Evaluation plan: specify metrics, test conditions, uncertainty and acceptance criteria before results are known.
7. Reproducibility: record code versions, parameters, random seeds, software and hardware.
8. Safety and ethics: assess failure modes, privacy, cybersecurity and human oversight.
9. Limitations: state where the results should not be used.
10. Deployment recommendation: distinguish laboratory evidence, pilot evidence and operational readiness.

Research integrity requires reporting negative results and failed experiments. Selective reporting of favourable test cases can produce unsafe confidence. Data containing household consumption, location or equipment identifiers may constitute personal or commercially sensitive information; collection and retention should be minimised and access controlled.

10. Contemporary Developments

Current engineering practice is moving toward foundation models for code, documentation and multimodal analysis; physics-informed and hybrid ML; federated learning across distributed assets; edge AI; synthetic data; explainable fault diagnosis; digital-thread integration and cyber-physical security. These developments offer efficiency but do not remove the need for validation.

Physics-informed learning incorporates governing relationships or constraints into training. It can improve plausibility where data are limited, but incorrect physical assumptions can bias results. Federated learning can reduce centralised data sharing, yet model updates may still leak information and require secure aggregation. Generative AI can assist with coding and design exploration, but generated content must be reviewed, tested and attributed appropriately. It should never be treated as an authoritative source of circuit ratings, protection settings or safety procedures without verification.

Challenges, Limitations and Governance

The principal limitation of AI in electrical engineering is not algorithmic capability but mismatch between training evidence and deployment reality. Sensor changes, ageing, new operating modes, extreme weather and adversarial inputs can invalidate a model.

Governance should address:

• responsibility for decisions and override authority;
• traceability of data, models and software versions;
• validation before deployment and after significant changes;
• monitoring for drift and abnormal confidence;
• cybersecurity and privacy;
• accessibility and fairness where users are affected;
• intellectual property and licensing;
• compliance with applicable electrical-safety, data-protection and AI regulations.

A risk-based approach is appropriate. A recommendation system for maintenance has different assurance needs from an autonomous protection relay. The higher the potential harm, the stronger the requirements for independent safeguards, testing, explainability, human oversight and formal change control.

Best-Practice Checklist

• Begin with an engineering problem, not an algorithm.
• Establish a defensible baseline.
• Use units, timestamps and physical limits consistently.
• Separate training, validation and final testing.
• Prevent temporal, identity and label leakage.
• Report performance by operating regime, not only as one average.
• Include uncertainty, failure modes and a fallback.
• Keep safety-critical protection independent.
• Version data, code, models and configurations.
• Test cybersecurity and monitor post-deployment drift.
• State limitations and define the conditions under which conclusions apply.

Chapter Summary

AI and digital engineering extend electrical engineering by connecting physical models, measurements, computation and lifecycle decisions. Machine learning supports prediction, classification, estimation and optimisation, but its reliability depends on representative data, correct evaluation and awareness of distribution shift. MAE, RMSE, precision and recall provide useful but incomplete evidence.

Digital twins combine models, live data and lifecycle information for monitoring and prediction. Automated design requires explicit objectives, constraints and sensitivity analysis. Cloud-connected IoT architectures improve coordination but introduce cybersecurity, latency and privacy risks. Smart grids, power electronics, autonomous systems and electrification provide important application domains, yet safety-critical functions require independent safeguards.

A strong dissertation defines its engineering question, establishes a baseline, documents data quality, uses an appropriate test strategy, reports uncertainty and limitations, and distinguishes experimental promise from deployment readiness. Responsible AI in electrical engineering is therefore a systems-engineering discipline rather than merely an exercise in selecting a sophisticated algorithm.

Key Terms

Artificial intelligence: Computational methods that perform tasks associated with perception, reasoning, prediction or decision-making.

Machine learning: Methods that infer patterns or relationships from data.

Supervised learning: Learning from examples with known labels or target values.

Unsupervised learning: Learning structure from data without supplied labels.

Reinforcement learning: Learning actions through interaction and reward signals.

Data leakage: Use of information in model development that would not be available at decision time.

Distribution shift: Difference between deployment data and training data.

Digital engineering: Lifecycle integration of models, data, software, simulation and physical systems.

Digital twin: A connected digital representation of a physical asset or process.

Edge computing: Processing data near the source rather than exclusively in a remote cloud.

Internet of Things: Networked physical devices capable of sensing, communicating or acting.

Explainability: The degree to which model behaviour can be understood by people.

Uncertainty: Lack of complete knowledge about measurements, parameters, predictions or outcomes.

Surrogate model: An approximate model used to replace an expensive simulation or experiment.

Pareto front: Set of non-dominated solutions representing trade-offs among objectives.

Concept drift: Change over time in the relationship between inputs and outputs.

State of charge: Estimated remaining charge of a battery relative to a defined capacity reference.

Hardware-in-the-loop: Testing in which real hardware interacts with a real-time simulated environment.

Cyber-physical system: A system combining computation, communication and physical processes.

Federated learning: Collaborative model training in which data remain distributed across participating sites.

Review Questions

Short Answer Questions

1. Distinguish verification from validation of an engineering model.
2. Why can accuracy be misleading for a rare-fault classifier?
3. Define data leakage and give one electrical-engineering example.
4. What distinguishes a digital twin from a static CAD model?
5. State two cybersecurity controls suitable for an IoT-enabled energy system.

Long Answer Questions

1. Explain how a student should design a reliable ML methodology for a time-series electrical-engineering dissertation.
2. Discuss the benefits and limitations of digital twins for battery or power-electronic systems.
3. Explain how AI can support smart-grid operation while respecting physical network constraints.
4. Compare cloud and edge computing for control, monitoring and maintenance applications.
5. Discuss why explainability, uncertainty and human oversight are important in safety-relevant electrical systems.

Analytical / Application Questions

1. A fault classifier detects 45 of 60 faults and generates 30 false alarms. Calculate recall and precision, and explain whether the system is suitable for autonomous protection.
2. Propose a chronological training, validation and test strategy for forecasting photovoltaic output. Identify two possible leakage pathways.
3. Design a high-level digital-engineering architecture for an electric-vehicle charging station. Specify sensing, edge control, cloud analytics, cybersecurity and fallback functions.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes a digital twin from a static CAD model?

A. A digital twin is always a three-dimensional drawing.

B. A digital twin is connected to physical-system data and supports lifecycle decisions.

C. A digital twin cannot contain mathematical models.

D. A digital twin is used only during manufacturing.

Correct Answer: B. A digital twin is connected to physical-system data and supports lifecycle decisions.

Question 2

Which practice most directly prevents temporal data leakage in a forecasting project?

A. Randomly shuffle all observations before splitting.

B. Use future observations to improve feature scaling.

C. Split the data chronologically and test on a later period.

D. Remove all observations containing missing values.

Correct Answer: C. Split the data chronologically and test on a later period.

Question 3

A classifier has TP = 80, TN = 900, FP = 20 and FN = 10. What is its precision?

A. 80%

B. 88.9%

C. 90%

D. 97.8%

Correct Answer: C. 90%

Question 4

Which metric is especially important when missed faults are more harmful than false alarms?

A. Recall

B. Mean absolute error

C. Specific energy

D. Coefficient of variation

Correct Answer: A. Recall

Question 5

What is the principal purpose of a validation dataset?

A. To fit the final model parameters only

B. To choose models and hyperparameters before final testing

C. To replace physical measurements

D. To guarantee zero deployment error

Correct Answer: B. To choose models and hyperparameters before final testing

Question 6

Which is an appropriate role for edge computing in a power-electronic system?

A. Performing latency-sensitive local processing near the equipment

B. Eliminating the need for sensors

C. Guaranteeing that cloud communication never fails

D. Replacing all independent protection functions

Correct Answer: A. Performing latency-sensitive local processing near the equipment

Question 7

For measured values and predictions, RMSE differs from MAE because RMSE:

A. Has no units

B. Penalises large errors more strongly

C. Can only be used for classification

D. Is always smaller than MAE

Correct Answer: B. Penalises large errors more strongly

Question 8

Which feature is most characteristic of responsible deployment of an AI maintenance system?

A. Removing all human review

B. Using only laboratory data

C. Maintaining a safe fallback and monitoring model drift

D. Optimising only the training score

Correct Answer: C. Maintaining a safe fallback and monitoring model drift

Question 9

In constrained automated design, a Pareto front represents:

A. Only designs that violate every constraint

B. Non-dominated trade-offs among objectives

C. A list of random initial guesses

D. The physical wiring diagram of a converter

Correct Answer: B. Non-dominated trade-offs among objectives

Question 10

Why should an AI model normally not replace certified protection in a safety-critical power system?

A. AI models cannot process numerical data

B. Protection never requires measurements

C. Model uncertainty and failure modes require independent safeguards

D. Protection systems operate only in the cloud

Correct Answer: C. Model uncertainty and failure modes require independent safeguards