电力系统预测主要包括负荷预测、出力预测以及健康状态预测等。通过负荷预测，可以优化电力生产规划，从而更好地实现电能的精细化分配；通过出力预测，可以有效提升新能源电力消纳能力，实现电能的充分及合理利用；通过电力设备健康状态预测，可以及时发现设备运行隐患，从而进一步保障电力系统平稳安全运行。深度学习凭借其卓越的特征分析和预测能力，被广泛应用于电力系统运行及维护。本文首先归纳介绍了电力系统预测深度学习模型的特点、适用场景；其次，梳理了深度学习在面向民用及工业场景负荷预测、光伏及风电出力预测、机械及非机械设备健康状态预测中的应用前沿；最后，对深度学习在电力系统预测中所面临的关键问题、发展趋势进行了总结和展望。 Abstract: Power system is one of the largest and complex artificial engineering in the modern society. With the development of intelligence, digitization and long-distance technology, a large number of multi-source, multi-state and heterogeneous operational data have emerged. As a new trend direction of machine learning, deep learning has shown potential in data feature extraction and pattern recognition. Because of its excellent ability in data analysis and prediction, it is widely used in power system, which has a significant impact on optimizing power production planning, improving power production efficiency and energy utilization, and ensuring the smooth operation of the system influence. Based on massive quantities of data and by means of deep learning, it can better fit the nonlinear relationship between the factors affecting the subsequent operational state of the system, so as to further improve the prediction accuracy. Power system prediction includes load forecasting, new energy power prediction and state-of-health prediction. Power production planning can be optimized using load forecasting; thus, electrical energy can be finely dispatched. The capacity of new energy power consumption is improved through power prediction to reasonably use electrical energy. Potential equipment hazards can be timely found using power equipment health state prediction, thereby ensuring safe and smooth operation. First, in this paper, the characteristics and applicable scenarios of typical deep learning models are introduced, among them, deep belief network and stacked auto encoder belong to stack structure, so the structure is flexible and easy to expand, which is suitable for the modeling and feature extraction of unrelated data type; convolutional neural network shares convolution kernel internally to reduce the number of network parameters and is good at processing high-dimensional data type; recurrent neural network has feedforward and feedback connections, so it is suitable for processing sequence data with pre and post dependence. Second, the application frontiers of predictive power systems based on deep learning are reviewed, which include civil and industrial scenarios, photovoltaic and wind power, mechanical and non-mechanical equipment health state prediction. Finally, facing the challenges of power system in energy efficient allocation, high proportion of new energy power consumption, highly stable operation of power equipment and so on, the key problems and future development trends are presented.