Abstract: ‍ ‍Cell-free radio access network （CF-RAN） breaks the traditional cellular networking method and forms a radio access network architecture with a scalable network node size. Based on the traditional cell-free massive MIMO （CF-mMIMO） systems， CF-RAN divides the physical layer functions， including remote radio unit （RRU）， edge distributed unit （EDU）， user-centered distributed unit （UCDU） and central processing unit （CPU）， which enable CF-mMIMO systems to achieve a trade-off between complexity and performance in cooperative transmission. In this network architecture， RRUs are responsible for completing the transmission and reception of radio frequency signals and digital-to-analog/analog-to-digital conversion； EDUs are responsible for completing channel estimation， multi-user/multi-data stream detection， multi-user/multi-data stream precoding， and reciprocity calibration of uplink and downlink channels in baseband signal processing； UCDUs are responsible for combining the received uplink data belonging to the same user sent by multiple EDUs， distributing multiple downlink data streams to the corresponding EDU， and channel estimation required for calibration between RRUs； and CPU is responsible for determining the data flow correspondence between EDUs and UCDUs. As a novel cell-free network architecture， a theoretical analysis of its system-level performance must be conducted， especially performance analysis that simultaneously considers the randomness of channel fading and the randomness of node spatial positions. This is the major research motivation of this study. We used a scalable full-pilot zero-forcing （FZF） system combining/precoding in uplink/downlink and considered the impact of channel estimation error and pilot contamination under the assumption of independent Rayleigh fading channels. We gave closed-form expressions of the uplink/downlink achievable signal-to-interference-noise ratio （SINR） of the CF-RAN system according to random matrix theory. Considering the spatial location randomness of user equipments （UE） and RRUs， we modeled the locations of UEs and RRUs as two independent binomial point processes （BPP） within a limited area， according to the distribution of various coherence distances between the UEs and RRUs. We derived the expression of the user downlink rate coverage probability of our scalable CF-RAN system， and the accuracy of our theoretical results was verified through Monte Carlo simulation. In simulation， we used genetic algorithm （GA）-based interleaving deployment， k-means algorithm-based clustering deployment， and random deployment on the EDU-RRU association scheme. According to the numerical analysis results， the scalable CF-RAN architecture achieved a flexible trade-off between centralized processing and distributed processing. The system performance was better than the fully distributed architecture and was close to the fully centralized architecture， while being scalable. In addition， compared with the other two node deployment schemes， the interleaved deployment based on GA achieved the best performance， while the clustering deployment based on the k-means algorithm caused the largest performance loss. Finally， according to the stochastic geometric analysis results， the user downlink coverage probability of scalable CF-RAN decreased as the number of EDUs and UEs increased. This research on the performance analysis of the scalable CF-RAN architecture provides some theoretical basis for the future development of user-centered network architecture and gives a new direction for cell-free research. However， this study has some shortcomings that need to be further improved. For example， after obtaining the closed-form expression of the uplink SINR， the uplink performance was not analyzed based on stochastic geometry theory.