Contextual intelligence for the energy system.

01 System context

The relationships are part of the data.

A meter reading, an equipment record and a market schedule can describe different aspects of the same asset. Their meaning depends on where that asset sits, what it connects to and when each piece of information applies.

ArayNex’s starting point is a unified representation of the physical and economic system. By contextual intelligence, we mean interpreting observations and potential actions in the context of the system as a whole. A weather change can affect generation, loading and available flexibility; an existing market commitment can determine how much of that flexibility remains usable.

One model means a consistent system context. Specialized models can work within it while referring to the same assets, time horizons and assumptions.

Structure

Assets, electrical connections, locations, equipment parameters and operating limits.

State

Measurements and estimates of voltage, flows, asset condition and available energy, tied to a point in time.

Conditions

Weather, generation and demand trajectories, planned outages and the uncertainty around them.

Commitments

Procurement positions, operating schedules, service obligations and the market rules that apply to an action.

Graph representations offer a way to encode assets as entities and connections as relationships. Learning over that structure can preserve dependencies that would be lost by treating every measurement as an independent input. A useful representation also distinguishes an electrical connection from a contractual relationship: both matter, but they impose different constraints. [1]

This context is intended to carry across networks, renewable generation, storage, computing, industry and mobility. The asset models and obligations differ by setting; the need to connect physical capability with economic choices remains.

02 Observations and state

Establish the system state from incomplete evidence.

Network records establish an initial structure. Telemetry describes behavior over time. Field imagery can provide evidence about asset identity, location or condition. Bringing these sources together requires matching them to the same equipment and distinguishing what was measured, inferred or confirmed in the field.

Keep the evidence attached.

A useful observation carries its source, timestamp, units, location and uncertainty. The time when an event occurred may differ from the time its record arrived. An equipment inspection can change an assumption about the asset without replacing a current sensor reading. Conflicting records need reconciliation, rather than silent overwriting.

Account for what is not observed.

State estimation relates available measurements to the network that produced them. Weighted least squares is an established engineering baseline: measurements contribute according to their uncertainty, and residuals help identify inconsistencies. It provides a meaningful comparison for learned approaches. [2]

Incomplete measurements can leave uncertainty about how the system is actually operating. Prior estimates and physical models can help reduce this uncertainty, but cannot replace direct observations. The representation should retain that uncertainty and identify when another observation or field check is needed.

03 Physics and learning

Learn within the physical principles that govern the system.

Connectivity, power balance and equipment dynamics provide information that a time series alone cannot supply. Our research direction brings those relationships into learning and evaluates learned estimates against engineering models.

Physics-informed neural networks (PINNs) incorporate governing equations into training alongside observed data. Depending on the formulation, a training objective can account for measurement error, equation residuals, and initial or boundary conditions. The original PINN work establishes this approach for differential equations. [3]

Match the model to the question.

Network state

Power-flow equations relate power injections, voltages and network flows at a given operating point. For unbalanced distribution networks, these relationships are modeled explicitly for each phase.

Asset dynamics

State changes depend on the asset: stored energy, temperature or electromechanical dynamics evolve across different timescales.

Learned approximation

A surrogate can estimate a state or response within a defined domain. Its inputs, training conditions and physical residuals remain part of the evaluation.

PINNs have been used to model rotor angle and frequency dynamics and estimate uncertain parameters such as inertia and damping. Each formulation is evaluated within a defined system and range of operating conditions. [4]

Models are compared with numerical time-domain solvers on accuracy, computational cost and flexibility across operating conditions. A model’s practical value depends on its training and retraining costs, inference speed and ability to generalize across operating conditions. [5]

Physical loss terms introduce their own optimization challenges. [6] Model outputs are evaluated against numerical baselines, checked against the governing equations and tested beyond the conditions represented in training. These checks assess accuracy and physical feasibility independently of training loss.

04 Scenarios and uncertainty

Evaluate futures that can occur together.

Weather influences both supply and demand. Renewable output, temperature-sensitive load and equipment availability can change together. Independent forecasts can miss the combinations that put the greatest pressure on the system.

Probabilistic forecasting represents a range of outcomes. For an operating plan, that range needs to extend across relevant locations and time periods. Wind forecasts describe possible generation trajectories across locations and time, while probabilistic demand forecasts capture uncertainty in electricity consumption. [7] [8]

Preserve sequence and dependence.

Storage availability depends on what happened in earlier intervals. A demand peak coinciding with low renewable output creates a different requirement from the same two events occurring apart. Network conditions, asset states and existing commitments provide a common context for evaluating these trajectories and the dependencies that shape the system’s response.

A forecast distribution and a stress test answer different questions. The first describes estimated probabilities; the second examines a specified condition, such as an outage. Keeping that distinction explicit prevents an assumed scenario from being mistaken for a calibrated likelihood.

Forecast evaluation should consider calibration and how informative the predicted ranges are. Proper scoring rules, including the continuous ranked probability score, provide tools for this assessment. Results also need examination by horizon, season and operating regime, so average scores do not conceal important failures. [9] For joint forecasts, evaluation also checks whether relationships across locations and time are preserved. [10]

05 Operations and markets

A useful choice must be physically possible.

Buying energy does not increase a connection’s transfer limit. Discharging storage changes what is available later. Reserving capacity for one commitment can make it unavailable for another.

Constrained optimization makes these relationships explicit. An objective expresses the cost or service outcome being considered; constraints define which responses are allowed. AC optimal power flow is one established formulation, combining power balance with network, voltage and generator limits. [11]

Physical feasibility

Balance, network loading, voltage limits, equipment availability and site connection limits.

Across time

State of charge, conversion losses, ramping, maintenance windows and future service obligations.

Economic exposure

Purchases and sales, deviation costs, tariffs, contractual positions and the applicable market rules.

Single-period feasibility is only part of an operating plan. Multi-period formulations connect generation and storage choices across time; open frameworks such as PyPSA combine these linked decisions with power-flow equations and asset constraints in a single optimization model. The appropriate physical approximation depends on the question, and any simplification needs to remain explicit.

Separate speed from feasibility.

A learned approximation may accelerate evaluation or provide a starting point for an optimizer. Its output still needs checking against the applicable equations and limits. In neural AC optimal power flow, physical constraints shape model training, while worst-case violation analysis provides a separate check on the feasibility of the resulting solutions. [13]

Judge forecasts by the choices they change.

An error near a binding constraint can change an operating decision more than a larger error under unconstrained conditions. Decision-focused learning studies this relationship between prediction and downstream optimization. It motivates assessing forecast accuracy and decision quality separately. [14]

ArayNex’s intended output is a response that can be examined: feasible alternatives, the constraints that determine them, the assumptions behind their cost and the commitments they create. Recommendations should retain operator review and decision authority, with their validity tied to the system conditions used to evaluate them.

06 Learning from outcomes

Update the understanding behind the next response.

Measurements, field findings and operating outcomes provide new context. Comparing them with expectations can reveal a forecast error, an incorrect equipment parameter, a topology change or an action that was never carried out.

A changed system

A replaced asset or confirmed switch change updates the physical representation and the assumptions of subsequent studies. Residual-based state estimation can help identify errors in the assumed network topology. [15]

A persistent error

A forecast bias or changing error distribution prompts investigation of data, features and calibration before deciding what to retrain. [16]

A different action

A recommendation that was modified or not executed must be distinguished from a model that predicted the executed action incorrectly.

Our design principle is to retain the observation, context version, recommendation and recorded action together. A prediction about one network configuration should not be assessed against another without accounting for the change.

An update then becomes something that can be tested. Replay it against historical periods using the information available at each point, compare it with the previous version, and examine whether any improvement survives changes in season, topology or operating conditions. This keeps contextual learning connected to evidence. [17]

07 Evaluation

Measure the foundation and the decisions it informs.

The foundation needs assessment both as individual models and as a connected system. A faster estimate is useful only if its error is understood. A lower-cost schedule is meaningful only under the constraints and implementation costs that would apply in practice.

Representation

Asset identity, connectivity and time alignment against trusted records; performance with missing, stale and conflicting observations.

State estimation

Voltage and flow errors against held-out measurements or an engineering baseline, reported by condition and network configuration.

Uncertainty

Calibration, interval coverage and forecast scores across horizons, with chronological separation between training and test periods.

Feasibility

Equation residuals and constraint violations, with units, tolerances, unsuccessful solves and any correction of model outputs reported.

Decision quality

Realized cost and service outcomes against a baseline with the same information, accounting for losses, deviations and implementation costs.

Computation

Data preparation, inference, optimization and verification time together; training and recalibration costs reported separately.

Compare under the same conditions.

A fair comparison fixes the information available at the time, the feasible actions and the operating constraints. It also asks which part of the foundation adds value: the network representation, physical information, uncertainty treatment or decision objective. Testing components separately helps identify the source of a gain or failure.

PGLib-OPF supplies standardized AC optimal-power-flow cases and formulations for reproducible algorithm comparisons. RTS-GMLC provides a complementary reference system for studies involving operating chronology and reliability. Their different scopes matter; neither substitutes for validation in the intended operating environment. [18] [19]

This perspective describes ArayNex’s technical direction and evaluation principles.