• The heat integration of the reference Kalina cycle is studied. • Increasing the internal heat recovery can improve the thermal performance. • The improvement in COP can be as high as 8.5%. • A thermal relation embedded artificial neural network is proposed. • The generation performance can be increased by at least 25% at small training sets. This paper presents an efficient ammonia-water power cycle based on the basic Kalina cycle. The heat exchanger network analysis (HEN) is applied to identify all the possible heat-saving configurations. There are 13 different cycles proposed, and each is analyzed and optimized under the same working condition. The results show that there exists a saturation region between the cycle complexity and thermal performance. The thermal efficiency of the optimum cycle has been increased by 8.5% compared to that of the reference Kalina cycle under the heat source temperature of 550℃ and cooling temperature of 24.5℃. The artificial neural network (ANN) surrogate model is also provided to alternatively determine the maximum thermal efficiency under the condition that there is sufficient training data available. The results are compared with those obtained from the thermal model within 0.06% to 1.2% differences. The common practices of ANN in the thermal systems are treated as black-box. Ignorance of thermal principles can lead to a misleading result when using the black-box model, especially for the small size of the training set. In this paper, a thermally constrained ANN architecture is proposed by considering the first law of thermodynamics. The results indicate that, as compared to the black-box model, the generalization performance can be increased at least by 25% with the optimum network configuration of 15-6-3-3 for 50 instances in the training set.