OPTIMIZATION OF RE-INJECTION IN LOW TEMPERATURE GEOTHERMAL RESERVOIRS USING NEURAL NETWORK AND KRIGING PROXIES
采出水回注进行压力保持是地热田管理中的常见做法。回注井的位置选择和注入速度对地热储层工程师来说是一个具有挑战性的课题。这类问题的优化目标通常是找到地热回注井位置的一个或多个组合，以最小成本和最小焓降最大限度地提高产量和压力保持。尽管可能有无数个回注井组合，但通常会预先选定一个网格作为井位，然后进行搜索，以找到满足生产目标的最具时间效益或成本效益的井位。
通常，使用模拟器开发代表性解决方案的知识库。然后训练并测试用于预测选定结果的人工神经网络。在下一步，将生成这些井的井组合和注入速度，以预测给定数量注入井的结果。另一方面，代表性解决方案的知识库可能用被克里金生成一个优化曲面，然后用于选择新的最佳搜索方向。本研究中，对比了神经网络代理方法和克里金代理方法快速评估的土耳其克孜勒卡哈姆低温地热田回注井位置。结果表明，神经网络代理方法比克里格代理方法更快、更准确。观察到克里金代理优化方法的精度取决于变异函数分析的精度。此外，克里格代理优化可能不会导致全局最优。
ABSTRACT
Re-injection of produced geothermal water for pressure support is a common practice in geothermal field management. The location selection of the reinjection well and the rate of injection is a challenging subject for geothermal reservoir engineers. The goal of optimization for this type of problem is usually to find one or more combinations of geothermal re-injection well locations that will maximize the production and the pressure support at minimum cost and minimum enthalpy decrease. Although the number of well combinations is potentially infinite, it has been customary to prespecify a grid of potentially good well locations and then formulate the search to locate the most time- or cost-effective subset of those locations that meets production goals. Typically, a knowledge base of representative solutions is developed using a simulator. Then an artificial neural network to predict selected outcomes is trained and tested. In the next step well combinations and injection rates of these wells to predict outcomes with a given number of injection wells are generated. On the other hand, knowledge base of representative solutions may be kriged to generate an optimization surface which then be used to select new optimal search directions. In this study, neural network proxy and kriging proxies for fast reinjection location evaluations are compared using low temperature Kizilcahamam, Turkey geothermal field. The results show that neural network proxy method is faster and more accurate then kriging proxy. It is observed that accuracy of kriging proxy optimization method depends on accuracy of variogram analysis. Moreover, kriging proxy optimization may not result in global optimum.
