A key barrier in the adoption of Wireless Sensor Networks (WSNs) is achieving long-lived and robust real-life deployments. Issues include: reducing the impact of transmission loss, node failure detection, accommodating multiple sensor modalities, and the energy requirement of the WSN network stack. In systems where radio transmissions are the largest energy consumer on a node, it follows that reducing the number of transmissions will, in turn, extend node lifetime. Research in this area has led to the development of the Dual Prediction Scheme (DPS). However, the design of specific DPS algorithms in the literature have not typically considered issues arising in real world deployments. Therefore, this thesis proposes solutions to enable DPSs to function in robust and long-lived real-world WSN deployments. To exemplify the proposed solutions, Cogent-House, an end-to-end open-source home environmental and energy monitoring system, is considered as a case study. Cogent-House was deployed in 37 homes generating 235 evaluation data traces, each spanning periods of two weeks to a year. DPSs presented within the literature are often lacking in the ability to handle several aspects of real world deployments. To address issues in real-life deployments this thesis proposes a novel generalised framework, named Generalised Dual Prediction Scheme (G-DPS). G-DPS provides: i) a multi-modal approach, ii) an acknowledgement scheme, iii) heartbeat messages, and iv) a method to calculate reconstructed data yield. G-DPS’s multi-modal approach allows multiple sensor’s readings to be combined into a single model, compared to single-modal which uses multiple instances of a DPS. Considering a node sensing temperature, humidity and CO2, the multi-modal approach transmissions are reduced by up to 27%, signal reconstruction accuracy is improved by up to 65%, and the energy requirement of nodes is reduced by 15% compared to single-modal DPS. In a lossy network use of acknowledgements improves signal reconstruction accuracy by up to 2x and increases the data yield of the system up to 7x, when compared to an acknowledgement-less scheme, with only up to a 1.13x increase in energy consumption. Heartbeat messages allow the detection of faulty nodes, and yet do not significantly impact the energy requirement of functioning nodes. Implementing DPS algorithms within the G-DPS framework enables robust deployments, as well as easier comparison of performance between differing approaches. DPSs focus on modelling sensed signals, allowing accurate reconstruction of the signal from fewer transmissions. Although transmission scan be reduced in this way, considerable savings are also possible at the application level. Given the information needs of a specific application, raw sensor measurement data is often highly compressible. This thesis proposes the Bare Necessities (BN) algorithm, which exploits on-node analytics by transforming data to information closer to the data source (the sensing device). This approach is evaluated in the context of a household monitoring application that reports the percentage of time a room of the home spends in various environmental conditions. BN can reduce the number of packets transmitted to the sink by 7000x compared to a sense-and-send approach. To support the implementation of the above solutions in achieving long lifetimes, this thesis explores the impact of the network stack on the energy consumption of low transmission sensor nodes. Considering a DPS achieving a 20x transmission reduction, the energy reduction of anode is only 1.3x when using the TinyOS network stack. This thesis proposes the Backbone Collection Tree Protocol (B-CTP), a networking approach utilising a persistent backbone network of powered nodes. B-CTP coupled with Linear Spanish Inquisition Protocol (L-SIP) decreases the energy requirement for sensing nodes by 13.4x compared to sense-and-send nodes using the TinyOS network stack. When B-CTP is coupled with BN an energy reduction of 14.1x is achieved. Finally, this thesis proposes a quadratic spline reconstruction method which improves signal reconstruction accuracy by 1.3x compared to commonly used linear interpolation or model prediction based reconstruction approaches. Incorporating sequence numbers into the quadratic spline method allows up to 5 hours of accurate signal imputation during transmission failure. In summary, the techniques presented in this thesis enable WSNs to be long-lived and robust in real-life deployments. Furthermore, the underlying approaches can be applied to existing techniques and implemented for a wide variety of applications.