Predicting forecasts for next 12 months using Box-Jenkins I am building a Box-Jenkins model in Excel using solver. The model is AR(2).
The data that I have contains trend and seasonality both. 
I know how to remove seasonality using seasonal indexes and add it back to the forecast. 
But, how do I handle trend? If I remove trend from the data, how should I add it back to the forecast?
Also, is the excel solver best way to find the AR parameters?
 A: If you are at all familiar with R (if you're building time series models, you should be), check out the forecast package. It's designed to choose parameters for Arima as well as exponential smoothing models, and uses a solid methodology to do so.  It will probably get you a lot farther than what you are building in excel, especially because it will also allow you to explore exponential smoothing models.  The two functions you are interested in are 'auto.arima' and 'ets.'
/Edit: auto.arima can also be used to fit ARMAX models, which (if properly specified) can solve many of the problems identified by IrishStat.
A: The time series are usually decomposed into 3 parts, trend, seasonality and irregular. (The link gives 4 parts, but cyclical and seasonality are usually lumped together). Strictly speaking ARIMA type of models are only used for irregular part and by their design these model do not incorporate any trend (I am assuming that trend is some function which varies in time). So if you simply want to estimate AR(2) model no software will estimate the trend for you, since if it did, it would not be fitting AR(2) model. 
To forecast the trend you will first need to create some sort of model and test it, and only after you are confident that your model truly estimates the trend, then you can use it to forecast the trend. Without such model any forecasting is impossible. Sadly the majority of time series textbooks do not stress this when talking about forecasting. 
A: Your approach suggests initially adjusting in a deterministic manner the impact of seasonality. This approach may or may not be applicable as the impact of seasonality may be auto-projective in form. The best way to answer this question is to evaluate alternative final models for adequacy in terms of separating the observed observations to signal and noise. There are a number of possible pitfalls awaiting you . One of them " does the series have one or more trends AND/OR one or more level shifts " ? Another possible issue "does the series have a constant set of monthly indicators or have some months had a statistically significant change in the their effects ? In terms of a seasonal ARIMA model this question translates to " have the model parameters changed over time " ?. My experience with Excel Solver has not been very positive.
A: As mentioned, Use R, not excel.
My understanding of this process you are asking for.
Say you have a data set with a linear trend.  Let's assume that trend to be Y = 3t+1, also assume you have 15 data points.
Use that model and find the residuals from that.  Fit your time series model to these residuals.  To forecast, use the 'predict' function in R. Get the next point.  Let's assume the model tells you the next error will be -2.
(If you wanted to predict the next point, this would be the 16th data point.)  Take Y= 3*16+1 = 49,,, and now add in the -2 from the time series prediction.  Your forecast is now :49+(-2) = 47.
