كتابة تعليق

Summary The authors forecast daily oil production for three wells on the Norwegian Continental Shelf: PF14 and PF12 from Volve, and E1H from Norne. They compare four models (LSTM, BiLSTM, GRU and XGBoost) that use bottom-hole pressure, water rate and gas rate as inputs. They chose these inputs based on inflow performance ideas. They use conformal prediction to add 95% uncertainty bands. Each model is trained on 80% of the history, tested on the remaining 20%, and then checked on a blind forecast of the next 10–12 months. The LSTM does best on PF14, and on E1H it beats the reservoir simulator’s own forecast.

Strengths

• The data is public (Volve and Norne), so other people can repeat the work.

• The blind test uses future data, not random samples, which matches how forecasts are used in the field.

• Giving uncertainty bands instead of just one number is useful for engineers.

• Comparing against the reservoir simulator is a practical benchmark. However, the numbers are only given for E1H; please give the same kind of table for PF14 and PF12.

Major comments

1. The 95% claim does not match the results. The abstract and conclusion say the method guarantees 95% coverage. But Section 2.4 reports only 70.34% for the PF14 LSTM, reports a different kind of number for E1H (7.4% of points outside the band), and gives nothing for PF12. Please report coverage the same way for every well and model, correct the abstract and conclusion, and explain why coverage is so low.

2. The same data is used to calibrate and to test. Section 2.4 says the bands are calibrated on the 20% test data, and coverage is then measured on that same data. The method’s guarantee only holds if these are separate. Also, daily production values are strongly linked to the days before them (the authors note this in Section 2.3), which breaks a key assumption of the method. Please use three separate parts (training, calibration and blind forecast), and consider conformal methods designed for time series.

3. The forecast uses the simulator’s future inputs. During the blind forecast, water rate, gas rate and pressure come from the reservoir simulator, not from measurements. For E1H, simulated pressure also replaces missing field data. So the model is predicting oil rate with the simulator’s help rather than on its own. Please state this clearly, show how errors in the simulated inputs affect the oil forecast, and test cases that use only inputs an operator would know in advance, such as planned pressure or choke settings.

4. Some table values look wrong.

   • (a) The XGBoost numbers in Table 2 (PF12) are exactly the same as in Table 1 (PF14), for example test MAE 34.437 and forecast MAE 87.742. This looks like a copy error.

   • (b) In several places the bias equals the MAE (for example BiLSTM on PF12: bias −281.845, MAE 281.845). That can only happen if every prediction is off in the same direction. Please check these.

   • (c) Section 3.1 praises BiLSTM on PF12, but Table 2 shows it has the second-worst forecast MAE (281.8) and sMAPE (192.5%), after GRU.

5. Two metrics are explained the wrong way round, and there are no simple baselines. By Equation (4), a positive bias means the model predicts too low, but the text says too high. For directional accuracy (PDA), higher should be better, but Section 3.1 says lower is better. Please also compare against simple methods, such as a decline curve (Arps) and “repeat the last value,” to show the deep learning models add something. Finally, the results come from only three wells, and the E1H test is one well, not real cross-validation, so please tone down the general claims.

6. The productivity index is described but never used. Equations (1)–(3) define the productivity index and the Wiggins IPR, but the model never calculates these. Reservoir pressure is fixed at 340 bar, and the raw rates and pressures go straight into the models. This is choosing inputs based on physics, not engineering new features. Please change the wording, or actually compute these features and show whether they help.

Minor comments

• Inputs: Section 2.1 lists oil rate, water rate and pressure, while Section 2.4 lists water rate, gas rate and pressure. Please make these match, and say whether past oil rates are also used as inputs.

• Please add a table of all model settings: lookback window, units, layers, batch size, learning rate and tree depth.

• Please say how much data was filled in by imputation, especially for the PF12 pressure data.

• On E1H the bias is positive in both periods (+53.8 and +135.3), so the model predicts too low throughout. Please discuss why.

• Please add a statement on whether the code and data are available.

• Typos: “Performance Index” should be “Productivity Index,” “CPI” should be “ICP,” and “Figures figs. 6 and 8” should be fixed.

Recommendation Major revision. This is a useful study with open data and a practical comparison against simulation. But the calibration needs to be fixed, the 95% claim has to match the results, the tables need correcting, and simple baselines should be added.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they used generative AI to come up with new ideas for their review.