Skip to main content
Research Q&A
66 main questions · 539+ sub-questions · Linked to thesis evidence
Problem → Framework → Results → Policy
↑↓ Navigate, Esc Clear filters
B2Layer 5 / 10hard~5 min read
What makes purely data-driven forecasting models ineffective for long-horizon energy planning?
Extrapolation breakdown (8.2% energy balance violations), no physical constraints, blindness to structural changes (35% forecast errors).
Key Takeaway
Data-driven models fail on long horizons due to extrapolation breakdown and lack of physical constraints.
Detailed Explanation
Data-driven models like the LSTM, ARIMA and Prophet suffer from three major failures in long-horizon energy forecasting in the period 2024–2050. Firstly, (1) Extrapolation breakdown Models trained on historical data from 2000-2023, characterized by incremental renewable growth and stable fossil capacity, malfunction when extrapolated to 2024–2050 requiring a trajectory of exponential renewable growth and aggressive phasing-out of fossils. The result is physically implausible forecasts (capacity factors > 1.0 and negative generation) and energy balance violations of 8.2%. (2) Not encoding physical constraints Pure neural networks don't possess the ability to encode domain knowledge allowing them to violate first principles like energy conservation or basic design. The following examples show it violates important modelling restrictions like (i) capacity factor bound [0,1] (ii) energy balance and (iii) monotonicity of cumulative capacity. (3) Blindness to structural changes Whatever the past policy regimes, the history of capacity deployment shows that they are not able to foresee the new regimes because the key policy variables are not part of their dataset. The policy regimes include (i) coal subsidy which allowed capacity models to forecast growing coal capacity, (ii) limited renewable deployment, but models cannot anticipate regime shifts (carbon pricing, coal phase-out, renewable acceleration) without explicit policy variables, causing forecast errors of 35% for technologies experiencing policy-driven acceleration.
Why It Matters
Research Graph — Related Questions
B2 What makes purely data-driven forecastin…
Sub-questions (10)
B2.1Layer 5 / 10hard~2 min read
What is extrapolation breakdown?
Models trained on 2000-2023 fail when extrapolated to 2024-2050.
Key Takeaway
Extrapolation violates physics.
Detailed Explanation
Data-driven models malfunction when extrapolated to trajectories requiring exponential renewable growth and fossil phase-out.
Why It Matters
Research significance: ENSM motivation.
Energy policy relevance: Forecast reliability.
Methodological contribution: S-curve bias.
B2.2Layer 5 / 10hard~2 min read
What are energy balance violations?
8.2% violations in LSTM; capacity factor >1, negative generation.
Key Takeaway
Pure NN violates conservation.
Detailed Explanation
LSTM produces 8.2% energy balance violations, capacity factors >1, and negative generation.
Why It Matters
Research significance: Physical constraint need.
Energy policy relevance: Unreliable forecasts.
Methodological contribution: ENSM constraints.
B2.3Layer 5 / 10hard~2 min read
What is capacity factor bound [0,1]?
Capacity factor must be between 0 and 1.
Key Takeaway
NN can violate this.
Detailed Explanation
Capacity factor—generation/capacity—must be in [0,1]. Pure neural nets can violate this.
Why It Matters
Research significance: Constraint design.
Energy policy relevance: Physical plausibility.
Methodological contribution: ENSM layer.
B2.4Layer 5 / 10hard~2 min read
What is monotonicity of cumulative capacity?
Cumulative capacity cannot decrease.
Key Takeaway
NN can violate monotonicity.
Detailed Explanation
Cumulative capacity is monotonically non-decreasing. Neural nets can violate this.
Why It Matters
Research significance: Constraint design.
Energy policy relevance: Realistic trajectories.
Methodological contribution: ENSM.
B2.5Layer 5 / 10hard~2 min read
What is blindness to structural changes?
Models cannot anticipate regime shifts.
Key Takeaway
Policy variables missing.
Detailed Explanation
Models trained on past regimes cannot anticipate carbon pricing, coal phase-out, or renewable acceleration.
Why It Matters
Research significance: Policy variable need.
Energy policy relevance: Regime shifts.
Methodological contribution: Input design.
B2.6Layer 5 / 10hard~2 min read
Why do LSTM/ARIMA/Prophet fail?
No physical constraints, extrapolation breakdown.
Key Takeaway
Physics-informed needed.
Detailed Explanation
LSTM, ARIMA, Prophet lack physical constraints and fail under extrapolation to new regimes.
Why It Matters
Research significance: Baseline limitation.
Energy policy relevance: Model choice.
Methodological contribution: ENSM vs baselines.
B2.7Layer 5 / 10hard~2 min read
What is S-curve inductive bias?
Sigmoid/Gompertz encode diffusion theory.
Key Takeaway
Limits to plausible trajectories.
Detailed Explanation
S-curve inductive bias embeds technology diffusion theory, limiting predictions to physically plausible trajectories.
Why It Matters
Research significance: ENSM design.
Energy policy relevance: Forecast quality.
Methodological contribution: C2.
B2.8Layer 5 / 10hard~2 min read
What are policy variables in forecasting?
Carbon price, subsidies, phase-out targets.
Key Takeaway
Missing in pure data-driven.
Detailed Explanation
Policy variables—carbon price, renewable subsidies, phase-out targets—enable anticipation of regime shifts.
Why It Matters
Research significance: Input design.
Energy policy relevance: Policy-relevant forecasts.
Methodological contribution: ENSM inputs.
B2.9Layer 5 / 10hard~1 min read
What is the 35% forecast error?
Technologies with policy-driven acceleration.
Key Takeaway
Regime shift impact.
Detailed Explanation
Forecast errors reach 35% for technologies experiencing policy-driven acceleration.
Why It Matters
Research significance: Quantifies failure.
Energy policy relevance: Error magnitude.
Methodological contribution: Benchmark.
B2.10Layer 5 / 10hard~2 min read
How does ENSM address these failures?
S-curve bias, hard constraints, Monte Carlo dropout.
Key Takeaway
Physics-informed solution.
Detailed Explanation
ENSM uses S-curve bias, hard physical constraints, and Monte Carlo dropout to reduce violations from 8.2% to <0.1%.
Why It Matters
Research significance: Solution validation.
Energy policy relevance: Reliable forecasts.
Methodological contribution: C2.