HomeFeaturesExamplesTutorialsHelpTheoryVerificationPricingContact
Theory manual

17. Thermal response test evaluation

A thermal response test (TRT) injects a heat rate Q(t) into a test borehole through the circulating fluid and records the inlet and outlet temperatures. The evaluation estimates the ground conductivity λ and the effective borehole resistance Rb* (optionally C, T0 and a Darcy velocity) and writes them into the design project. Reference standards: VDI 4640 Part 5 (2020), ISO 17628:2015 and the ASHRAE Handbook. The program leads through the evaluation in six steps (import, setup, quality, evaluation, comparison, apply) and writes a TRT report with the contents of VDI 4640 Part 5.

17.1 Data, cleaning and quality checks

The records (time, Tin, Tout, flow, optional measured power and optional mean fluid temperature, for example a DTS depth mean) are never modified; cleaning only sets flags. The mean fluid temperature is Tf=(Tin+Tout)/2 unless given; the power is measured, or

Q=ρcp(T¯)V˙(Tin−Tout)
(17.1)

with the fluid properties at the record's mean temperature; a sample's value holds over the interval that ends at it. The heating ends at the last sample above half the typical heating power. Spikes are found by a Hampel identifier (±7 samples, 4 × 1.4826 × MAD and at least 0.1 K or 5 %) on complete windows only, so genuine steps are never flagged; spikes and user-excluded intervals are left out of the fits, but their heat stays in the load history. Table 23 lists the quality checks; a failed check is reported as a failure of the test data, and every limit is editable.

Table 23. Quality checks of the TRT evaluation.
CheckDefaultIf violatedSource
Heating duration≥ 48 hFailVDI 4640-5, ISO 17628; ASHRAE ≥ 36–48 h
Power: sd / mean (10-min means, first hour excluded)≤ 1.5 %FailASHRAE
Power: largest deviation / mean≤ 10 %FailASHRAE
Specific heat rate Q¯/H30–80 W/mWarningASHRAE 50–80 W/m; VDI 4640-5
Flow: sd / mean≤ 5 %Warninggood practice
Reynolds number at the mean flow≥ 2300Warningturbulent flow recommended
Days between grouting and test≥ 5 dWarningASHRAE
T0 against Ts+qgeo(D+H/2)/λ≤ 2 KWarningplausibility
Energy balance: measured power against (17.1)≤ 10 %Failsensor consistency
Data gaps (spacing > 3 × median)longest ≤ 1 h, share ≤ 5 %Warninggood practice
Heating interruptions (Q < 50 % of the median)noneWarningM1 invalid across them
Spikes removed≤ 1 % of the samplesWarning–

17.2 Forward model

For a heat rate per metre q(t)=Q(t)/H that changes by Δqk at tk the mean fluid temperature is

Tf(t)=T0+Rb*q~(t)+12πλ∑kΔqkG(t−tk)
(17.2)

with G the response of one borehole at its wall: the infinite line source (4.20) (ILS, the default and reference model); the hollow infinite cylinder source (CHS, G=2πG(Fo) of (4.23), tabulated with a relative error below 10−8), in which the heat enters the ground at rb and the borehole itself stores no heat; the finite line source (FLS) of one segment of length H at depth D, which is the line source with end effects only; and the moving line source of M3. q~ is q after a first-order lag of time constant τb (default 1 h), which represents the delay of fluid and grout: q~k=qk+(q~k−1−qk)e−Δt/τb. The sum is a causal convolution on a uniform grid (at most 16 384 steps) evaluated with the FFT, so one model evaluation costs O(NlogN).

Why the line source is the reference. At the wall the hollow cylinder differs from the line source by

GCHS−GILS≈ln(4Fo)+1−γ4Fo−18Fo(Fo≫1)
(17.3)

which decays only like ln(Fo)/Fo: the slope of GCHS against lnt is 11 % below that of the line source at Fo = 10, 8 % at 20, 4 % at 50 and 1 % at 250 (Figure 15, exact quadrature; the expansion (17.3) gives 13, 8, 4 and 1 %). Over the usual windows (Fo ≈ 5–250) a CHS fit therefore reads λ 4–9 % lower than an ILS fit of the same data, whatever the lag, the window, T0 and the power history. A real borehole is not hollow: its core (fluid, pipes, grout) stores heat. For a perfectly conducting core of heat capacity S per metre (Carslaw & Jaeger 1959, §13.7; Blackwell 1954) the slope relative to the line source is:

Table 24. Slope over the evaluation window relative to the line source for a borehole core with heat capacity (independent quadrature of the kernels).
S/(πrb2C)0 (hollow = CHS)0.511.52
Beier sandbox, Fo 10–530.9340.9610.9891.0181.047
synthetic example, Fo 7–530.9270.9570.9871.0181.050
Darmstadt logger, Fo 5–2480.9530.9720.9921.0121.032

Usual boreholes have S/(πrb2C) ≈ 0.8–1.5 (water 4.2, pipes 1.8, grout 1.5–3.9 MJ/(m³ K) against 2.0–2.6 in the ground), where the line source is within ±2 % of the core model and the hollow cylinder 5–7 % off. The line source represents the borehole's own heat capacity implicitly, the CHS neglects it; this is why the ILS is the reference model.

Slope of the hollow-cylinder response against ln t relative to the line source, from an independent quadrature of the kernels, with the Fourier-number ranges of two evaluation windows.
Figure 15. Slope of the hollow-cylinder response against ln t relative to the line source, from an independent quadrature of the kernels, with the Fourier-number ranges of two evaluation windows.

17.3 Methods M1, M2, M3 and M5

M1, line-source slope (Mogensen 1983). For t≫rb2/α and constant power, (17.2) with the ILS becomes Tf=klnt+m, and

λ=Q¯4πHk,Rb*=m−T0q¯−ln(4α/rb2)−γ4πλ
(17.4)

k and m come from least squares over the evaluation window (default from max(10 h, 5rb2/α) to the heating end); their covariance gives the standard errors by the delta method and 95 % intervals with the Student t quantile. The sequential evaluation repeats M1 for 40 window ends and 40 window starts (exact individual regressions by prefix sums); when λ rises by more than 5 % over the last half of the window ends, M3 is suggested (a drifting power has the same signature). M1 needs a constant power: a 12 % drift biases it by 5–7 %.

M2, parameter estimation with superposition. λ, Rb* (optionally C and T0) minimise ∑(Tf,model−Tf,measured)2 over the window (and the recovery when included) with (17.2) and the chosen G, by Levenberg–Marquardt with box bounds through a logistic transformation (λ 0.1–20 W/(m K), Rb* −0.2–2 m K/W, C 0.2–8 MJ/(m³ K), T0 ± 50 K), started from M1. The covariance at the solution

Cov=s2(𝐉T𝐉)−1,s2=SSRn−m
(17.5)

gives standard errors, intervals and the correlation matrix. They assume independent residuals; real TRT residuals are strongly correlated (lag-1 autocorrelation r1 = 0.93–0.99), so the intervals are optimistic and the Monte Carlo analysis covers the systematic input errors. The effective number of independent observations (the AR(1) variance-inflation factor) is reported with every fit:

neff=n1−r11+r1,r1∈[0, 0.999]
(17.6)

Identifiability (every method). A fit is not identifiable, and is then shown as a sensitivity but never recommended automatically, when a parameter is at its search bound; when λ and its level parameter (Rb*, for M5 T0) have |ρ|>0.999 (they are always strongly correlated, ρ ≈ 0.98–0.99); when any other pair has |ρ|>0.90 (C trades against λ through α, T0 against Rb*); when the 95 % interval of λ is wider than ±20 %; and, for M3, when the advection effect vT2tend/(4α) is below 0.05 or the interval of u includes 0.

M3, moving infinite line source (Diao et al. 2004; Molina-Giraldo et al. 2011; Wagner et al. 2013). With groundwater flow the mean of the MILS temperature over the wall is

G=12I0(b)∫rb2/(4αt)∞exp(−ψ−b24ψ)dψψ,b=vTrb2α
(17.7)

It is the ILS for u=0 and tends to I0(b)K0(b). M3 estimates λ, Rb* and u from three starting values of u. The flow direction cannot be identified from one borehole, and a bounded test domain also looks like advection.

M5, recovery. The classic form after a heating period th with mean power Q¯ assumes a constant power, the logarithmic kernel and the line source:

Tf(t)−T0=Q¯4πλHlntt−th
(17.8)

BoreTherm superposes the whole heating history with the exact kernel G of the model chosen for M2:

Tf(t)=T0+S(t;α)2πλ,S(t;α)=∑kΔqkG(α(t−tk))
(17.9)

which is linear in 1/(2πλ) and T0 for a given α (α=λ/C is updated by fixed-point iteration). The recovery has no Rb* term. On cylinder-source data with 24 h of recovery the classic form read +15.6 %; the decomposition in Table 25 shows that the kernel is the cause (the recovery samples lie at Fo ≈ 7–17 since the stop, where the kernels differ most), so M5 is only as good as the kernel it uses.

Table 25. Error of the M5 conductivity on the synthetic example (noise-free, samples 10–24 h after the stop; independent evaluation of the kernels).
Data, powerClassic (17.8)Log kernel + power historySuperposition, ILS kernelSuperposition, CHS kernel
ILS, constant+2.6 %+2.6 %0.0 %−13.5 %
ILS, swing + stop+1.6 %+2.6 %0.0 %−13.5 %
CHS, constant+16.8 %+16.8 %+14.1 %0.0 %
CHS, swing + stop+15.6 %+16.8 %+14.2 %0.0 %

17.4 Undisturbed temperature and uncertainty

T0 is the mean fluid temperature of the circulation before the heating, a manual value, or the mean of a measured temperature profile over [D,D+H]; without it M2 estimates T0 and M1 gives no Rb*. The Monte Carlo analysis draws N Latin-hypercube realisations (the sampler of Chapter 10) of normal input errors (power or flow, the offsets of the two sensors, H, rb, C, T0; defaults 2 %, 1 %, 0.05 K, 0.5 m, 2 mm, 10 %, 0.2 K) and re-runs M1 and the recommended fit on each perturbed data set, in parallel and deterministic for a seed; the result is P5/P50/P95 of λ and Rb*.

17.5 Recommended value

A fit that is not identifiable by its own diagnostics is a sensitivity: shown, marked, never chosen automatically. The rule:

  1. User choice: the preferred fit, if it succeeded (also when not identifiable; the reason says so).
  2. M2 forward model: among the identifiable M2 fits the line source is the reference; the FLS or the CHS replaces it only when it fits significantly better by the Bayesian information criterion with the effective number of observations:
ΔBIC=nefflnSSRrefSSRalt−Δmlnneff≥10
(17.10)

with neff from (17.6) (the larger r1 of the two fits) and Δm the number of additional parameters (0 between M2 models, 1 for M3). With r1 = 0.93–0.99 the 2000–4000 observations of a TRT count as 10–150 independent ones; a plain RMSE comparison or an F-test on n would call every small improvement significant. 10 is “very strong” evidence on the scale of Kass & Raftery (1995).

  1. M3 replaces the chosen M2 fit when it is identifiable (including |ρ(λ,u)|≤0.90) and (17.10) holds against it with Δm = 1.
  2. Otherwise the chosen M2 fit; otherwise M1 if identifiable; otherwise M5 if identifiable.
  3. Otherwise nothing is recommended, and applying a result needs an explicit choice.

The reason text names the rule that applied, with ΔBIC and neff, and the notes list the fits set aside. M5 is the independent check, not a candidate, while M2 succeeds.

17.6 Examples

Table 26. Evaluation of the two TRT examples with every method and model; the recommended fit in bold. Deviations against the generating values (synthetic: λ = 2.60 W/(m K), Rb* = 0.085 m K/W) and against the independently measured values (sandbox: λ = 2.88 W/(m K), Rb = 0.165 m K/W).
DataMethodλ [W/(m K)]ΔλRb* [m K/W]ΔRb*RMSE [K]neff
Synthetic example (cylinder-source data)M1 line-source slope2.7154.4 %0.09177.9 %0.6362
M2 superposition (ILS)2.8298.8 %0.095612.4 %0.09749
M2 superposition (CHS)2.599−0.1 %0.08500.0 %0.0171498
M2 superposition (FLS)2.8208.5 %0.095512.3 %0.09749
M3 moving line source (sensitivity)2.8298.8 %0.095612.4 %0.09749
M5 recovery (CHS)2.593−0.3 %––0.014421
Beier et al. (2011) sandboxM1 line-source slope2.9231.5 %0.1579−4.3 %0.03629
M2 superposition (ILS)2.831−1.7 %0.1548−6.2 %0.05314
M2 superposition (CHS)2.639−8.4 %0.1457−11.7 %0.06018
M2 superposition (FLS)2.746−4.6 %0.1536−6.9 %0.05115
M3 moving line source (sensitivity)2.140−25.7 %0.1372−16.8 %0.03333

On the synthetic example, whose data were generated with the cylinder source, the rule finds the CHS (ΔBIC 159 over the ILS with 46 effective observations) and recovers λ and Rb* within 0.1 %; M5 with the CHS kernel reads −0.3 %. On the Beier sandbox (a 18.3 m borehole in a 1.8 m box, heater power 1056 W) the line source is kept: the FLS fits slightly better but not significantly (ΔBIC 0.9), and the CHS worse. M3 has the lowest RMSE but λ and u are correlated (ρ = −0.96): the bounded box looks like advection, so M3 is a sensitivity only. The Monte Carlo analysis (200 samples) gives P5–P95 of λ = 2.68–3.00 W/(m K) for the recommended fit. Figure 16 shows the fits and the sequential evaluation. On the Darmstadt MD-BTES test (750 m coaxial borehole; Seib et al. 2025) the recommended M2 (ILS) gives 2.277 W/(m K) and 0.0547 m K/W against the paper's 2.28 / 0.05; the data set is not distributed with BoreTherm, and only these numbers are quoted.

Beier et al. (2011) sandbox TRT: (a) measured mean fluid temperature and the M2 fits with the three forward models over the default window; (b) sequential M1 evaluation, λ against the window end and the window start.
Figure 16. Beier et al. (2011) sandbox TRT: (a) measured mean fluid temperature and the M2 fits with the three forward models over the default window; (b) sequential M1 evaluation, λ against the window end and the window start.

17.7 Applying the results to the design

Applying a result writes the chosen fit into a project and lists every change. λ goes to the homogeneous ground, to every layer, or to one layer i; λTRT is the effective value over the test length, so when the effective value is to be matched, layer i gets

λi=λTRT−∑j≠iwjλjwi
(17.11)

with the length shares wj of (3.3), so that the project's effective conductivity over the test length equals λTRT. C is written only when it was estimated. The undisturbed temperature T0 shifts the surface temperature so that the undisturbed temperature over the test length equals T0. The effective resistance Rb* is written either as a fixed value (classic style: it no longer varies with the design flow and length) or, recommended, as the grout conductivity kg that makes the multipole Rb* of the test borehole at the test flow and mean fluid temperature equal the TRT value; Rb* decreases monotonically with kg, so the root is bracketed on [0.05, 20] W/(m K) and found by the Illinois method in lnkg. The design then computes Rb* for its own flow and length. The Darcy velocity of M3 is stored on request in the ground's groundwater input, and the design simulation then uses it (Chapter 12).

Import. Delimited text files (CSV, TXT, Modelica tables) with detected delimiter and decimal comma, columns by letter, number or header, units per column, date-time formats (ISO, German, US), repeated time stamps averaged or spread, and presets for published data sets; wide DTS files (fibre positions × time stamps, several GB) are streamed one fibre position at a time (a 1.5 GB file takes about 10 s).