Application of multilayer perceptron neural network and seismic multiattribute transforms in reservoir characterization of Pennay field, Niger Delta

Abstract

Supervised multilayer perceptron neural network and seismic multiattribute transforms were applied to three-dimensional (3D) seismic and a suite of borehole log data set obtained from Pennay field, offshore Niger Delta with a view to predicting lateral continuity of hydrocarbon reservoir properties beyond well control. Four (4) hydrocarbon-bearing sands, namely, Pennay 1, 2, 3, and 4, were delineated from borehole log data. Four (4) horizons corresponding to near top of mapped hydrocarbon-bearing sands were used to produce time maps and then depth structural maps using appropriate checkshot data. Petrophysical analysis of the mapped reservoirs revealed that the area is characterized with hydrocarbon saturation ranging from 56 to 72%, water saturations between 27 and 44%, volume of shale between 7 and 20%, and porosity between 25 and 31%. Three major structure building faults (F2, F3, and F5 which are normal, listric concave in nature), two antithetic (F1 and F4), were identified. Structural closures identified as roll-over anticlines and displayed on the time and depth structure maps suggest probable hydrocarbon accumulation at the upthrown side of the fault F4. Seismic attribute results reveal two main characteristic patterns of high and low amplitude and frequency areas. There is an intermediate zone between the low and high amplitude and frequency that may be regarded as transition zone. Multilayer Perceptron Neural Networks (MLPNN) revealed how permeability, net-to-gross, porosity, volume of shale, and hydrocarbon saturation vary away from well control across the entire field. The supervised MLPNN-simulated volumes for petrophysical properties of interest have their uncertainties quantified and measure of accuracy in prediction and the predictive power in terms of root-mean-square error (RMSE). RMSE is the standard deviation of residual which is difference between the predicted values and observed values, i.e., input and output of the network. RSME ranges between 0 and 1 and values closer to zero (0) indicate a fit that is more useful for prediction and also very high confidence level for the resulting output. Permeability modelled at RMSE of 0.030 revealed some thief zones (channel with high absolute permeability) within and outside the areas with well concentration with average permeability of 635md. MLPNN-modelled map of net-to-gross (NTG) at RMSE of 0.0290 revealed that 72% of the reservoirs have a very high NTG (ratio of the volume of the sand in the reservoir to the total volume of the reservoir) with average NTG of 0.7184. Effective porosity was modelled at RMS error of 0.0053 with resulting average effective porosity of 0.295. The effective porosity slice on top of reservoirs revealed the lateral variation of effective porosity across the field with very high effective porosity areas coincide with the delineated regions of interests. Moreover, hydrocarbon saturation was also modelled at RMSE of 0.0282 with an average of 69.7% and volume of shale was modelled at RMSE of 0.028 and average volume of shale of 9%. A comparison of MLPNN slices of petrophysical properties on the mapped horizons revealed that a relatively higher NTG, low volume of shale, high hydrocarbon saturation, high permeability, and high effective porosity were observed in the regions of interest in the field. In conclusion, successful MLPNN prediction has been done for petrophysical properties at inter-well points and locations beyond well control. MLPNN-modelled maps revealed some bypassed sand channels and some thief zones within and outside the areas with well concentration that are not evident on the structural maps and attribute slices. The integration of the different analyses and results from the study has improved our understanding of mapped reservoirs and enhanced lateral prediction of its properties.