Federating Decline Curves Across Operators: Sharing the Map from Early Data to Decline Parameters

Oil and gas wells in shale, tight gas and coalbed methane fields are still forecast mostly with decline curve analysis: fitting a fixed formula, such as Arps' hyperbolic decline, to a well's past production and extrapolating it forward. This works poorly when it matters most. Early in a well's life the data are short and noisy, and the flow passes through several stages (fast linear flow out of the hydraulic fractures, then slower depletion of the reservoir boundary), which a single formula cannot follow. Machine learning forecasters do better, but they need many wells' worth of history. That history is split among competing operators who will not pool it. We propose a federated forecaster, meaning one trained jointly by several parties without any party sending its raw data to the others. What gets trained jointly is narrow: a small network that reads a new well's first few months of production plus its completion design and outputs the parameters of a multi-stage decline curve. Each operator keeps a private correction model on top. That correction is driven by whatever pressure and operational sensors that operator happens to have.

The bet is that the transferable knowledge across operators is "wells completed like this, which started like this, decline like this." It is not the full weights of a sequence model. If the shared part is restricted to that physically meaningful, low-dimensional mapping, two things should follow. A well with two or three months of history should be forecast nearly as well as by a model trained on everyone's pooled data. And the federation should suffer much less from operators' data looking different from one another than standard federated averaging of a full neural forecaster does.

Core Claim

Decline curve analysis fails at early time because a single well does not contain enough information to pin down its own decline parameters. That is a data-scarcity problem, and federation can address it without asking anyone to give up the parametric curve that engineers trust. We write each well's rate as

$$ q_i(t) = D\big(t;\ \theta_i\big)\,\exp\!\big(r_{\phi_k}(u_i(t))\big), \qquad \theta_i = g_\psi\big(s_i,\ q_i(1{:}\tau)\big), $$

where $q_i(t)$ is the rate of well $i$ at time $t$. $D$ is a two-segment decline: rate falling as $t^{-1/2}$ during fracture-dominated linear flow, then switching at a learned time $t_s$ to Arps hyperbolic decline with initial decline rate $d$ and exponent $b$. $\theta_i = (q_0, t_s, d, b)$ collects those parameters, with $q_0$ normalized by lateral length. $g_\psi$ is the shared encoder. Its inputs are the static completion and geology attributes $s_i$ and the first $\tau$ months of production. $r_{\phi_k}$ is operator $k$'s private residual network. Its input $u_i(t)$ is the operational signals available to that operator, such as casing pressure, choke changes or shut-ins. Only $\psi$ is federated. The residual is multiplicative and bounded, so the forecast stays a decline curve unless local operational data say otherwise. This takes the physics-constraint idea from P-C-L-A and makes the constraint structural rather than a penalty term.

What is new is where the sharing happens. Prior federated forecasters share whole models in weight space. Here the shared object is a map into decline-parameter space, whose outputs have the same physical meaning in every basin, even when rate scales and sensor sets differ.

Relation to Prior Work

Three papers motivate the problem and supply the evidence that plain decline curves break.

Review of Machine Learning Methods for Steady State Capacity and Transient Production Forecasting surveys why decline curve analysis and numerical simulation fail on multi-stage unconventional flow. It names federated learning as a future direction but implements nothing. We supply the concrete design and test.

Prediction of Coalbed Methane Production Using a Modified Machine Learning Methodology shows that Arps, Li-Horne, exponential and harmonic curves fail on volatile coalbed methane wells. An LSTM with attention reaches R² of 0.79, 0.92 and 0.56 on three wells, where a plain LSTM falls to −1.04 on the fluctuating well. It also finds that pressure and liquid-level signals matter more than geology. That finding is why our operational signals enter a private residual and not the shared encoder.

Machine Learning-Based Production Prediction Model and Its Application in Duvernay Formation uses Gaussian process regression on completion and geology attributes to predict six-month cumulative gas. It reports that its model does not carry over to other formations. This basin-specificity is the drift problem our parameter-space sharing targets.

Taken together, these papers and the coalbed methane paper above disagree about whether static geology or dynamic engineering signals drive forecasts. Our split assigns static attributes to the shared side and dynamic signals to the local side, and an ablation tests that split.

Two papers supply the method. Research on Intelligent Production Optimization of Low-Permeability Tight Gas Wells embeds Arps and SEPD decline laws into a CNN-LSTM-attention loss with weight 0.6, all on centralized data from four wells. For new wells with no history it falls back on nearest-neighbor matching (7.21% average relative error). We keep the decline physics but make it the output structure. Federated training then replaces nearest-neighbor matching for new wells. FedHealth shows that shared lower layers plus locally personalized upper layers help when clients differ. We reuse that split, but our shared part outputs physical parameters, not generic features.

Several federated papers constrain the design. Tackling Data Heterogeneity in Federated Time Series Forecasting (Fed-TREND) handles clients with mismatched sensors and dynamics by having the server build small synthetic training sets. It reports beating centralized training on six of eight benchmarks. It is our strongest generic federated competitor. FedSkeleton federates graph forecasting across parties that hold connected parts of a network. Inter-well interference across lease lines is a real effect we deliberately leave out of scope, so wells are treated as independent given their attributes. FTA-FTL is the only upstream federated study in the file. It classifies core images and loses under one point of accuracy to centralized training (90.68% versus 89.83% on uneven splits), which shows operators will tolerate federation for geological data. It does not touch time series. Production Forecasting Based on Attribute-Augmented Spatiotemporal Graph Convolutional Network shows that neighboring injectors matter. It cuts mean absolute error by 37.3% versus a single-well recurrent network. This is the main threat to our independence assumption in waterflooded or gas-injected fields, which is why we restrict the claim to primary-depletion unconventional wells.

Research Program

We will build the model in the equation above. The two-segment decline is implemented differentiably, with a smooth switch at $t_s$, so the encoder can be trained end to end on forecast error over the held-out future. Training alternates in rounds. Each operator updates $\psi$ and its own $\phi_k$ on local wells. The server averages only the $\psi$ updates, weighted by well count, under secure aggregation. Because $g_\psi$ is a small network with a few thousand parameters, communication per round is trivial, and compression schemes like those of Deep Anomaly Detection for Time-series Data in Industrial IoT are unnecessary.

The analysis has two parts. First, we measure how forecast error falls as the observed history length $\tau$ grows, under four regimes: local-only, federated, centralized, and plain curve fitting. The quantity of interest is the history length at which federated forecasting matches centralized accuracy. Second, we measure client drift directly. We take the spread of operators' local updates at the start of each round, expressed both in encoder output space and in weight space, and compare it against federated averaging of a full LSTM forecaster.

The data plan is honest about what is public. The primary benchmark is the Duvernay horizontal-well population used by Guo et al., drawn from Alberta regulatory records, provided the monthly histories behind their cumulative targets can be retrieved from the same public records. Clients are formed by partitioning wells by operating company. This gives a natural, uneven federated split and a centralized oracle on identical data. Because public records are monthly and lack pressure data, the private-residual component can only be tested where partner data exist. The daily-resolution fields in the file are the Sulige tight gas wells, the 21 Yangquan coalbed methane wells and the Middle East carbonate producers. All three are proprietary. Treating them as separate clients in a cross-basin federation requires data agreements, and we treat that as a second phase. If those agreements fail, the fallback is a set of simulated wells. These are generated with varied fracture and matrix properties and assigned per-client sensor subsets. This tests the mechanism but not field realism.

Baselines

The naive reference is per-well curve fitting. We fit Arps hyperbolic decline and, separately, SEPD to each well's first $\tau$ months by least squares and extrapolate. This is the incumbent industry practice and the thing the proposal must beat at short $\tau$. Beating it means lower error on 12- and 24-month cumulative production for $\tau \le 6$ months. At $\tau = 24$ months we expect it to be competitive, and we will say so if it is.

The strongest centralized reference on the public benchmark is the Gaussian process regression of Guo et al.. It reports R² of 0.80 and RMSE of 280.54 × 10⁴ m³ for six-month cumulative gas on a 129/30 well split. We rerun it on our splits with its eight attributes. We also train our own architecture centralized on pooled data as an upper bound. Federation counts as successful if it comes within 5% of centralized RMSE at $\tau = 3$ months. It counts as clearly beating local-only training if it reduces error by more than the spread across random operator splits.

The strongest federated competitors are federated averaging and FedProx applied to a full LSTM-attention forecaster in the style of the coalbed methane study, and Fed-TREND applied to the same backbone. For Fed-TREND, the file gives no reported result on Lacuna for production forecasting. Its reported advantage is beating centralized training on six of eight non-subsurface benchmarks, without per-dataset numbers in the file. All federated methods get the same number of communication rounds, the same local epochs, and the same hyperparameter search budget of 30 trials, tuned on a validation operator held out from testing.

For the daily-resolution fields, the reference is P-C-L-A trained locally per field. It reports average 30-day RMSE of 0.1618 × 10⁴ m³ on Sulige stable wells and a 24.41% RMSE reduction over a backpropagation network.

Three ablations isolate the key ingredient. The first is a federated type-curve model: each operator shares only summary statistics of fitted decline parameters, binned by completion attributes, and new wells take the pooled bin prior. The second is our model with the encoder replaced by a federated linear regression from attributes to decline parameters, with no early-production input. The third is our model with the local residual removed. A fourth ablation moves operational signals into the shared encoder, which tests the geology-versus-engineering split.

Why not just…?

Why not just fit Arps or SEPD curves locally on each well? They need no collaboration and leak nothing. Their failure is exactly at short history. The coalbed methane study shows classical decline models failing to follow volatile wells, and the tight gas study needed nearest-neighbor borrowing for new wells. The error-versus-$\tau$ curve in our experiments shows where, if anywhere, local fitting catches up.

Why not just share type curves, a pooled prior over decline parameters by completion bin? This is the real threat. It is already industry practice and could capture most of the benefit for long-horizon recovery. Our ablation runs exactly this through the same federation. If it matches the full model at $\tau \ge 3$ months, the honest conclusion is that operators should share binned decline statistics, not train networks, and we will report that. Our expectation is that the encoder wins only through its use of the first months of production, which a bin prior ignores. That is a testable and possibly small margin.

Why not just pool public regulatory data centrally and fine-tune locally? On the public Duvernay benchmark this is equivalent to our centralized oracle, so the federated gain there is meaningful only relative to local-only training. The case for federation rests on private daily operational data and unreported completion details, which public records lack. Only the partner-data phase can show whether those matter. If the residual ablation shows negligible gain, centralized pretraining on public data is the better recommendation.

Why not just federate a full neural forecaster with FedAvg or Fed-TREND? These are the direct comparison. Our hypothesis is that sharing in decline-parameter space reduces drift across operators with different rate scales and sensors. The drift measurement and the matched-budget comparison settle it. If Fed-TREND's synthetic-data calibration closes the gap, the parametric structure buys interpretability and physical plausibility but not accuracy. We will report the rate of forecasts that increase during undisturbed depletion as a separate metric so that trade-off is visible.

Decisive Checks

The proposal succeeds if three things hold. On operator-partitioned Duvernay data, the federated decline encoder at three months of history beats per-well Arps and SEPD fitting and local-only training on 12- and 24-month cumulative error. It lands within 5% of the centralized model. And it beats the federated type-curve ablation by more than split-to-split variance. The mechanism is supported if the spread of operators' updates is measurably smaller in encoder output space than for full-model federated averaging, and if that gap tracks the accuracy gap across splits. The proposal fails cleanly if the type-curve ablation matches it. In that case the field needs better-shared decline statistics, not federated networks, and the result is still worth publishing.