Does Renewable Energy Development Reduce Fossil Fuel Import Dependence?

Evidence from Taiwan's Energy Generation

Jun-Jie Chang — Department of Agricultural Economics, National Taiwan University College Student Research Scholarship, National Science and Technology Council (NSTC)

---

The headline

Over Taiwan's transition period (2016–2026), a 1% increase in renewable generation is associated with a 0.039% reduction in power-sector fossil fuel imports. Bootstrap 95% CI: [−0.075, −0.001].

That is fractional displacement, not substitution — less than one twenty-fifth of one percent. And the mechanism is identifiable: dispatchable renewables displace imports about 11.7× more strongly than variable solar and wind, whose effect is statistically indistinguishable from zero.

The puzzle in one picture — renewable generation quadruples after 2016 while power-sector fossil imports show no downward turn:

_{Figure F2. Deliberately drawn as separate panels on a shared time axis, not as a dual-axis chart. Two independent y-scales can be aligned arbitrarily, which manufactures apparent correlation; a paper arguing that renewables have not displaced fossil fuel should not open with a figure that fabricates the opposite impression.}

Taiwan is an unusually clean setting for this question. It is an isolated island grid — no interconnectors, so no cross-border electricity trade can absorb intermittency (Jacobson, 2021) — and it simultaneously ran two large policy shocks: a renewable build-out from 2016 and a nuclear phase-out completed in May 2025.

---

Why this is not already known

Taiwan imported 95.7% of its total energy supply in 2024. Reducing that dependence is the stated objective of the energy transition, and capacity targets are the policy instrument. But the link from capacity to reduced imports has never been tested for Taiwan.

_{Figure F1. Import dependence falls from 97.81% (2000) to 95.72% (2024) to 94.84% (2026:01) — about two percentage points across a quarter century. The 2024 annual mean reproduces the Energy Administration's published 95.8%, which cross-validates the series.}

The international literature does not settle it either. York (2012) finds that alternative energy largely adds to rather than displaces fossil fuel. Marques et al. (2018), covering ten European countries, find a substitution effect for hydropower and solar PV but not for wind, and stress that flexible, controllable capacity is what backs renewables up. Karlilar Pata & Balcilar (2024) report roughly 1.15 units of renewable capacity per unit of fossil capacity displaced across the OECD. None of these covers an island grid, which is precisely where intermittency should bite hardest.

There is also a trap worth naming, because the obvious approach falls into it. Taiwan's official "Imported Energy / Energy Supply (%)" series looks exactly like the dependent variable this paper wants — and it correlates with ln(RE) at −0.930. That number is not a finding. Renewables are domestic supply, so they sit in the denominator: the correlation is an accounting identity. Regressing that ratio on renewable generation produces a large, significant, entirely tautological coefficient. This study therefore measures imports in physical energy-equivalent terms and uses the official series only for descriptive figures and source cross-validation.

---

Data

Balanced monthly panel, 2000-01 to 2026-01 — 313 observations × 30 variables, no missing months, no NaNs.

Variable	Content	Source
FFI_power	Power-sector fossil imports, TJ — steam coal + sub-bituminous + LNG	MOEA E-STAT, countries-of-origin series
FFI_all	All fossil imports, TJ — adds crude oil and coking coal (robustness only)	as above
RE	Renewable generation, MWh — split into RE_disp (hydro + biomass + waste + geothermal) and RE_var (solar + wind)	MOEA E-STAT
EC	Electricity consumption, MWh, total and industrial	MOEA E-STAT
nuc_share	Nuclear generation ÷ electricity consumption	MOEA E-STAT
ER	NTD/USD, monthly average	Central Bank of the Republic of China (Taiwan)
PF	Divisia-weighted fuel price index (coal + LNG)	World Bank Pink Sheet

Three measurement choices carry most of the weight:

Imports are converted with the Energy Administration's official calorific-value table, not international rules of thumb. The three source files arrive in incompatible units (tonnes, thousand barrels, tonnes) and cannot simply be added. One item remains open: the official table gives LNG on a post-gasification volume basis while the import data are liquid mass, and the tonne→m³ factor is not published, so LNG retains the international convention of ~52 GJ/t.

_{Figure F3. The composition of RE changes completely within the sample: hydro falls from 71% of renewable generation in 2000 to 15% in 2025, while solar PV and wind rise from nothing to 75% combined. Pre-2010 ln(RE) is essentially a rainfall series — dispatchable, climate-driven, policy-irrelevant. Running one elasticity across 2000–2026 would have ln(RE) standing for two different economic mechanisms, which is the substantive reason the main sample begins in 2016. Smoothed to a 12-month mean, stated on the axis, because the figure carries a slow structural change rather than monthly variation.}

Crude oil and coking coal are excluded from the main specification. Crude oil alone is 38–55% of imported energy, but Taiwan barely burns it for electricity — it goes to refining, transport and petrochemicals, and coking coal goes to steel. Including them injects variance unrelated to generation, biasing the elasticity toward zero, where it would be misread as evidence for the additionality paradox. The de-trended, de-seasonalized residual correlation with ln(RE) makes the point: −0.010 for 2000–2015 versus −0.289 for 2016 onward.

Prices are international benchmarks, not Taiwan's own unit values. Taiwan's realized import prices are endogenous — they embed its own contract structure, procurement timing and freight, all correlated with its own load and dispatch decisions. Benchmark series are exogenous to a price-taker. The specific series (Coal, Australian; LNG, Japan; Crude oil, Dubai) were chosen against Taiwan's actual sourcing mix: Australia supplies 50.7% of coal, Australia + Qatar 56.4% of LNG, the Middle East 70.7% of crude. Henry Hub is explicitly rejected — gas markets are regionally segmented, and in 2022-09 Henry Hub read 7.76 against Japan LNG at 23.73.

_{Figure F4. The three benchmark series entering the Divisia index. The 2021–23 spike is the shock that makes a price control indispensable: without it, the exchange-rate coefficient would absorb variation that belongs to prices.}

---

Method

An ARDL bounds-testing framework (Pesaran, Shin & Smith, 2001), estimated as an unrestricted error correction model. ARDL is the right tool here because the regressors are a genuine mix of I(0) and I(1): ln(FFI_power) is I(1) in every sample (the bounds test's one hard prerequisite), while the renewable series are trend-stationary over the transition period.

$$ \begin{aligned} \Delta \ln FFI_t = \alpha_0 &+ \lambda_y \ln FFI_{t-1} + \sum_j \theta_j x_{j,t-1} \\ &+ \sum_i \phi_i \Delta \ln FFI_{t-i} + \sum_j \sum_i \delta_{ji} \Delta x_{j,t-i} + \psi D_t + \varepsilon_t \end{aligned} $$

with $x = {\ln RE,\ \ln EC,\ nuc_share,\ \ln ER,\ \ln PF}$, centred monthly dummies $D_t$, and the long-run elasticity recovered as $\eta_{RE} = -\theta_{RE}/\lambda_y$.

Choices that matter for credibility:

• Cointegration is confirmed by the three-test protocol of McNown et al. (2018) — overall $F$, the dependent-variable $t$, and the independent-variable $F$ — which rules out the degenerate cases a single $F$-test cannot detect. Critical values are simulated at the actual sample length, 100,000 replications, rather than borrowed from asymptotic tables.
• $\eta_{RE}$ is a ratio of estimates and its distribution is skewed, so inference leads with a 2,000-replication bootstrap CI and reports the delta-method interval alongside.
• Lag orders are AIC-selected within the candidate set that passes Ljung–Box(12), so no specification is bought at the price of autocorrelated residuals. Maximum lag 2; M1 selects p=2, all q=1 — 25 parameters on 119 observations.
• A structural break at 2022-11 (Chow F = 24.57, p = 2.0×10⁻¹⁰, matching the post-invasion energy shock) is absorbed by a dummy. Zivot–Andrews puts the exchange-rate break at 2022-02 independently.
• Monthly import data are lumpy — shipments, not economics: ACF(1) of Δln is −0.35, and it vanishes under quarterly aggregation (−0.01). Monthly frequency is retained for the observation count, with the quarterly specification (M6) as the check.

flowchart LR
    A["data/<br/>E-STAT · Central Bank<br/>World Bank Pink Sheet"] --> B["build_panel.py<br/><i>heat values, Divisia index</i>"]
    B --> C["panel_monthly_2000_2026.csv<br/><b>313 × 30</b>"]
    C --> D["run_specs.py<br/><i>ardl.py engine</i>"]
    D --> E["output/results/<br/>M1–M10, CUSUM, multipliers"]
    E --> F["make_outputs.py"]
    F --> G["output/tables/ T1–T7<br/>output/figures/ F1–F8"]
    G --> H["paper/main.tex<br/><i>no number typed by hand</i>"]

Loading

---

Findings

1 — Fractional displacement is confirmed. $\eta_{RE} \approx -0.039$: statistically supported, economically inelastic. Consistent in sign with York (2012) and Karlilar Pata & Balcilar (2024), and squarely in the range their frameworks predict.

2 — Intermittency is the mechanism. This is the paper's strongest contribution. Splitting renewables (M4): dispatchable hydro + biomass + waste + geothermal carry $\beta = -0.0373$; variable solar + wind carry $\beta = -0.0032$ (p = 0.83). Dispatchable displacement is 11.7× larger, and variable renewables display essentially none. This replicates Marques et al. (2018) and extends it to the island-grid case the literature had not covered — the setting where the result should be sharpest, and is.

3 — The displacement effect is recent. The 96-month rolling window (M10, 218 windows) shows $\eta_{RE}$ falling from +0.1896 at its 2010–11 peak — when renewables were essentially rainfall, and expansion ran alongside import growth in pure energy addition — to −0.0506 at the end of the sample, with the confidence band clearing zero only after roughly 2022. This is empirical justification for focusing on the transition period, not a data-availability excuse.

_{Figure F6. The sign reversal is the point. An elasticity that is positive in windows ending around 2010 and negative in windows ending after 2022 is not a stable structural parameter across the full period — it is a relationship that came into existence. Note also how the confidence band narrows as renewables grow: the early positive estimates rest on a period when there was barely any variation in renewable generation to identify from.}

Specification matrix

All ten specifications pass the full cointegration and diagnostic battery. All nine point estimates of $\eta_{RE}$ are negative.

	Variation	$\eta_{RE}$	p	Bootstrap 95% CI
M1	Main: FFI_power, ln RE, 2016+, Case III	−0.0391*	0.061	[−0.075, −0.001]
M2	FFI_all (adds crude oil + coking coal)	−0.0375	0.221	[−0.090, +0.016]
M3	Full sample from 2000-01	−0.0422*	0.075	[−0.086, +0.003]
M4	RE split: dispatchable vs. variable	−0.0032	0.827	[−0.032, +0.025]
M5	Drop nuc_share (collinearity check)	−0.0228	0.374	[−0.070, +0.023]
M6	Quarterly frequency (shipment lumpiness)	−0.0739**	0.016	[−0.123, −0.027]
M7	Fourier seasonality (order 3) vs. 11 dummies	−0.0403**	0.044	[−0.075, −0.003]
M8	Case V (unrestricted intercept and trend)	−0.0754**	0.010	[−0.127, −0.025]
M9	Fuel-specific prices vs. the Divisia index	−0.0377*	0.084	[−0.076, +0.001]
M10	96-month rolling window, 218 windows	+0.190 → −0.051	—	see finding 3

Bootstrap intervals exclude zero for M1, M6, M7, M8, M9. They include zero for M2 and M3, for M4 (variable renewables — which is the finding), and for M5.

Long-run coefficients, M1

$$ \ln FFI_{power} = -0.039 \ln RE + 0.555 \ln EC - 0.480, nuc_share - 0.145 \ln ER + 0.047 \ln PF $$

	β	SE	p	Reading
ln RE	−0.0391*	0.0209	0.061	the elasticity of interest
ln EC	+0.5551***	0.1888	0.003	inelastic but firmly significant
nuc_share	−0.4795***	0.1655	0.004	semi-elasticity; sign as expected
ln ER	−0.1452	0.1232	0.239	depreciation dampens imports, not significant
ln PF	+0.0466***	0.0116	<0.001	tiny magnitude, wrong sign — see below

n = 119 · 25 parameters · adjusted R² = 0.776 · $\lambda_y$ = −1.653

_{Figure F5. The cumulative response to a one-percent renewable generation shock is −0.0378 on impact and settles at −0.0391 by month 3. Adjustment is fast, and the convergence value reproduces the long-run coefficient from the table above through an entirely separate computation — a check on the estimator, not a restatement.}

Diagnostics

Every specification clears serial correlation, ARCH, normality, functional form and parameter stability. Selected p-values for M1: Ljung–Box(12) 0.139 · ARCH(12) 0.271 · Jarque–Bera 0.898 · Ramsey RESET 0.652 · CUSUM and CUSUMSQ both stable.

_{Figure F7. Brown–Durbin–Evans recursive-residual tests. Both statistics stay inside the 5% bands for the main specification, so the long-run relationship is not being held together by a single sub-period.}

The anticipated collinearity problem did not materialize: maximum VIF is 3.38, despite ln RE and nuc_share correlating at −0.767 (renewable expansion and nuclear retirement ran concurrently in opposite directions). Dropping nuc_share (M5) attenuates $\eta_{RE}$ from −0.039 toward zero — exactly the direction omitted-variable bias predicts — which is the argument for keeping it.

---

Stated limitations

Three results are reported rather than smoothed over, and the paper argues each one:

The adjustment speed exceeds unity — $\lambda_y$ = −1.653 for M1, between −1.05 and −1.69 across specifications. This is overshooting, not instability: |1 + λ| = 0.653 < 1, so the system still converges, oscillating as it does. It is what the shipment-lumpiness evidence predicts, and the quarterly specification gives λ = −1.363, so it is not an artefact of monthly frequency.

The fuel-price coefficient is positive (+0.047, p < 0.001). Literally, "fuel gets dearer, imports rise." Plausibly simultaneity — price and quantity both responding to demand — or long-term contract pricing decoupling the spot price from realized procurement cost. The magnitude is small enough not to threaten identification of $\eta_{RE}$, but it is stated openly.

ln EC is unstable across specifications — +0.555*** in M1, around +1.35 in the full sample, insignificant in M2. The M2 failure is informative rather than troubling: FFI_all contains crude oil and coking coal, which have little to do with electricity demand. That is the exclusion argument, arriving as evidence.

Two further constraints, deliberately accepted: no temperature control (seasonal dummies capture the average summer, not the unusually hot one), and generation rather than capacity as the renewable measure, which limits direct comparison with the capacity-based literature.

---

Repository

data/                   Raw sources, unmodified
  EC/  ER/  FFI/  HEATVALUE/  IND/  NUC/  PRICE/  RE/
scripts/
  build_panel.py        Merge + derive -> panel (heat values and sample window are top-level constants)
  ardl.py               UECM engine; hand-built design matrix, self-validated against statsmodels
  run_specs.py          Specification matrix M1-M10
  make_outputs.py       LaTeX tables + figures
output/
  panel_monthly_2000_2026.csv     313 x 30
  results/                        Unit roots, long/short-run coefficients, CUSUM, multipliers, summary.json
  figures/ F1-F8.png              300 dpi, print-ready
  tables/  T1-T7.tex              booktabs + threeparttable
paper/
  main.tex                        Full report; tables \input from ../output/tables/
  reference.bib                   28 entries, every one cited
C802/                   Original NSTC research proposal (folder named after the
                        NSTC C802 application form it contains)
EDA/                    Exploratory figures produced from the panel

Reproduce

python3 scripts/build_panel.py    # rebuild the panel (313 x 30)
python3 scripts/run_specs.py      # estimate M1-M10 -> output/results/
python3 scripts/make_outputs.py   # tables + figures -> output/{tables,figures}/
cd paper && latexmk -pdf main.tex

Python 3 · pandas 2.3.3 · openpyxl 3.1.5 · statsmodels 0.14.6 · matplotlib 3.10.8 · seaborn 0.13.2. LaTeX requires biblatex + biber (APA style).

---

References

Works cited above. The full bibliography — 28 entries, every one cited in the report — is paper/reference.bib.

Substitution and displacement

• Jacobson, M. Z. (2021). The cost of grid stability with 100% clean, renewable energy for all purposes when countries are isolated versus interconnected. Renewable Energy, 179, 1065–1075. doi:10.1016/j.renene.2021.07.115
• Karlilar Pata, S., & Balcilar, M. (2024). Decarbonizing energy: Evaluating fossil fuel displacement by renewables in OECD countries. Environmental Science and Pollution Research, 31(21), 31304–31313. doi:10.1007/s11356-024-33324-8
• Marques, A. C., Fuinhas, J. A., & Pereira, D. A. (2018). Have fossil fuels been substituted by renewables? An empirical assessment for 10 European countries. Energy Policy, 116, 257–265. doi:10.1016/j.enpol.2018.02.021
• York, R. (2012). Do alternative energy sources displace fossil fuels? Nature Climate Change, 2(6), 441–443. doi:10.1038/nclimate1451

Taiwan context

• Chuang, M. C., & Ma, H. W. (2013). An assessment of Taiwan's energy policy using multi-dimensional energy security indicators. Renewable and Sustainable Energy Reviews, 17, 301–311. doi:10.1016/j.rser.2012.09.034

Methodology

• Brown, R. L., Durbin, J., & Evans, J. M. (1975). Techniques for testing the constancy of regression relationships over time. Journal of the Royal Statistical Society: Series B, 37(2), 149–163. doi:10.1111/j.2517-6161.1975.tb01532.x
• Diewert, W. E. (1976). Exact and superlative index numbers. Journal of Econometrics, 4(2), 115–145. doi:10.1016/0304-4076(76)90009-9
• Efron, B. (1979). Bootstrap methods: Another look at the jackknife. The Annals of Statistics, 7(1), 1–26. doi:10.1214/aos/1176344552
• Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis of stationarity against the alternative of a unit root. Journal of Econometrics, 54(1–3), 159–178. doi:10.1016/0304-4076(92)90104-Y
• McNown, R., Sam, C. Y., & Goh, S. K. (2018). Bootstrapping the autoregressive distributed lag test for cointegration. Applied Economics, 50(13), 1509–1521. doi:10.1080/00036846.2017.1366643
• Narayan, P. K. (2005). The saving and investment nexus for China: Evidence from cointegration tests. Applied Economics, 37(17), 1979–1990. doi:10.1080/00036840500278103
• Pesaran, M. H., Shin, Y., & Smith, R. J. (2001). Bounds testing approaches to the analysis of level relationships. Journal of Applied Econometrics, 16(3), 289–326. doi:10.1002/jae.616
• Zivot, E., & Andrews, D. W. K. (1992). Further evidence on the great crash, the oil-price shock, and the unit-root hypothesis. Journal of Business & Economic Statistics, 10(3), 251–270. doi:10.1080/07350015.1992.10509904

Data sources

• Central Bank of the Republic of China (Taiwan). Monthly average exchange rate of the New Taiwan dollar against the US dollar (series BP01M01).
• Energy Administration, Ministry of Economic Affairs. Energy Statistics Query System (E-STAT) — monthly series on imports by country of origin, electricity generation, electricity consumption, energy indicators.
• Energy Administration, Ministry of Economic Affairs. Unit calorific values of energy products (effective from 2024).
• World Bank. Commodity Markets Outlook: Pink Sheet monthly historical prices — Coal (Australia), Liquefied natural gas (Japan), Crude oil (Dubai).

Software

• Seabold, S., & Perktold, J. (2010). statsmodels: Econometric and statistical modeling with Python. Proceedings of the 9th Python in Science Conference, 92–96.
• Waskom, M. L. (2021). seaborn: Statistical data visualization. Journal of Open Source Software, 6(60), 3021. doi:10.21105/joss.03021

---

_{Significance: * p<0.10, ** p<0.05, *** p<0.01. Data current through 2026-01.}