Despite advances in the understanding of fire dynamics over the past decades and despite the advances in computational capacity, our ability to predict the behaviour of fires in general and building fires in particular remains very limited. This thesis proposes and studies a method to use measurements of the real event in order to steer and accelerate fire simulations. This technology aims at providing forecasts of the fire development with a positive lead time, i.e. the forecast of future events is ready before those events take place. A simplified fire spread model is implemented, and sensor data are assimilated into the model in order to estimate the parameters that characterize the spread model and thus recover information lost by approximations. The assimilation process is posed as an inverse problem, which is solved minimizing a non linear cost function that measures the distance between sensor data and the forward model. In order to accelerate the optimization procedure, the 'tangent linear model' is implemented, i.e. the forward model is linearized around the initial guess of the governing parameters that are to be estimated, thus approximating the cost function by a quadratic function. The methodology was tested first with a simple two-zone forward model, and then with a coarse grid Computational Fluid Dynamics (CFD) fire model as forward model. Observations for the inverse modelling were generated using a fine grid CFD simulation in order to illustrate the methodology. A test case with observations from a real scale fire test is presented at the end of this document. In the two-zone model approach the spread rate, entrainment coefficient and gas transport time are the governing invariant parameters that are estimated. The parameters could be estimated correctly and the temperature and the height of the hot layer were reproduced satisfactorily. Moreover, the heat release rate and growth rate were estimated correctly with a positive lead time of up to 30 s. The results showed that the simple mass and heat balances and plume correlation of the zone model were enough to satisfactorily forecast the main features of the fire, and that positive lead times are possible. With the CFD forward model the growth rate, fuel mass loss rate and other parameters of a fire were estimated by assimilating measurements from the fire into the model. It was shown that with a field type forward model it is possible to estimate the growth rates of several different spread rates simultaneously. A coarse grid CFD model with very short computation times was used to assimilate measurements and it was shown that spatially resolved forecasts can be obtained in reasonable time, when combined with observations from the fire. The assimilation of observations from a real scale fire test into a coarse grid CFD model showed that the estimation of a fire growth parameter is possible in complicated scenarios in reasonable time, and that the resulting forecasts at localized level present good levels of accuracy. The proposed methodology is still subject to ongoing research. The limited capability of the forward model to represent the true fire has to be addressed with more detail, and the additional information that has to be provided in order to run the simulations has to be investigated. When using a CFD type forward model, additional to the detailed geometry, it is necessary to establish the location of the fire origin and the potential fuel load before starting the assimilation cycle. While the fire origin can be located easily (as a first approximation the location of the highest temperature reading can be used), the fuel load is potentially very variable and its exact distribution might be impractical to continually keep track of. It was however shown that for relatively small compartments the exact fuel distribution is not essential in order to produce an adequate forecast, and the fuel load could for example be established based on a statistical analysis of typical compartment layouts.