| Popis: |
We present an assessment on the predictability of the rate of penetration using artificial neural networks when only drilling records from an adjacent are at hand. The study was carried out with data from two Enhanced Geothermal System wells in South Korea. We compare five data-arrangement cases for neural network training. Differences in the specific values of parameters from the adjacent and the predicted well, such as the weight on bit, and the rotary speed influenced the prediction errors. Among the five cases, the highest error (85.3%) occurred for the case of cumulative data up to the point of prediction. However, some data-arrangements decrease the error, e.g. when employing the whole data (46.8%), or when using small data sections near prediction (case 4: 42.4%, and case 5: 39.7%). We suggest the combination of these last tree approaches when prediction ROP with limited data. |
