Residential users have huge flexibility potential, and the full exploiting and reasonable utilizing residential-side flexibility resources can help to improve the flexibility of the power grid. In this paper, an assessment method of residential-side flexibility resources based on non-intrusive load monitoring is proposed through using low-frequency power data and deep learning models. A power fluctuation-skipping event detection algorithm is used to realize the localization of appliance power events and power data acquisition. Time convolutional networks (TCNs) and gated recurrent units (GRUs) are combined to construct a TCN-BiGRU load recognition algorithm with the help of the data feature extraction capability of TCNs and the nonlinear fitting capability of GRUs to efficiently differentiate the electricity loads of different appliances. The load identification results are used to decompose the total power signal of users, establish the equipment state matrix, equipment probability matrix and equipment habitual use interval matrix, obtain the power consumption information of each appliance, analyze the energy consumption behavior of users, and obtain the detailed results of the flexibility resource assessment on the residential side. The practical effectiveness of the proposed method is verified by actual residential user data. The flexibility resource assessment results obtained based on the proposed method can provide auxiliary decision-making for residential demand-side response.