How to turn a time series problem into a cross-section problem A fairly frequent problem in the world of machine learning and predictions in general are the cases of cross-section and time series predictions.
A time series is represented basically by an observation over time measured in constant time spaces. For example the sales of one store per day,
where every remark is the sum of the store's sale in one day.
Another possibility would be a cross-section which is a photograph of a single moment of time (which may add several moments). For example, each observation can be a different store and each column the value of the sale in a period (Example: a day, a month or other aggregations).
Doubt is, how best to transform a time series problem (when we have observations over time) into a cross-section problem.
We can simply consider each day of sale as a cross-section even though it is not independent? the traditional algorithms of machine learning (thinking of predicting any target) would normally work in this case?
 A: Shameless plug for my paper:  Machine Learning Methods for Crop Yield Prediction and Climate Change Impact Assessment in Agriculture
First, reshape your data such that each observation is a unit-period.  I.e.: take it from wide to long format.  gather in tidyverse-speak.
If you were using classical statistics, and you had repeated observations of many units over time, you might use a model that looks like this:
$$
y_{it} = \alpha_i + \mathbf{X}_{it}\beta +\epsilon_{it}
$$
where the $\alpha$ is a fixed or random intercept (capturing unobserved cross-sectional heterogeneity), and the covariates $\mathbf{X}_{it}$ can include things that you think or know are important, like linear/quadratic/low-dimensional time trends that are additively separable but yet estimated simultaneously with everything else.  
If you knew from theory or whatever that this model was correctly specified up to a parameter vector, you'd just fit OLS or maybe ridge.  If you believed that the true data-generating process was nested in the above equation, you'd run LASSO.  
But if you believed that the model was useful but mis-specified, you might consider fitting
$$
y_{it} = \alpha_i + \mathbf{X}_{it}\beta + f(\mathbf{Z}_{it})+\epsilon_{it}
$$
where $f()$ is a neural net.  In principle, it could be any nonparametric algorithm, but a neural net is convenient because both the parametric and nonparametric part have gradients and can be optimized -- simultaneously -- by gradient descent.  (Other approaches to models like this include the backfitting algorithm, but I haven't had much success using the backfitting algorith where $f()$ is a random forest or xgboost).  
The addition of parametric structure to the neural net improves efficiency.  If $N\rightarrow \infty$, the net would probably figure out the time trends and individual heterogeneity, and any other parametric structure.  But in finite samples, it can really help.
While I've done this in pure R, it can also be done in modern deep learning libraries like Keras/Tensorflow, with a bit of hacking.
If you had a dynamic process, you could apply similar logic to augment an autoregressive model with a RNN or LSTM, though dynamic models have their own considerations that don't always apply in the panel context.
A: You can't transform a problem from time series to cross-sectional. Let's get that out of the way.
What you can do is to apply to time series problems the techniques that are typically associated with cross sectional analysis. For instance, a typical OLS problem is cross-sectional, because some of its assumptions are reasonable for many cross-sectional cases. Particularly, independence of error terms from regressors. 
Consider sales $S_i$ in stores $i$. You can often argue that in a model $S_i=X_i\beta+\varepsilon_i$, the errors $\varepsilon_i$ are not correlated with the predictors $X_i$. This is especially easy to argue with random sampling and experiments, where you randomly pick the order and subjects into your sample or control the values of regressors $X_i$.
Note, that when the sampling is not random or when the regressors are not controlled, even the cross-sectional regression becomes difficult. Such is the case of business forecasting. For instance, it is not always possible to control some predictors for the retails stores such as the outside weather. You get the weather that was that day, you can't set the outside temperature and humidity to your liking.
With time series the problems become even more difficult because the observations are ordered in time, they can't be random in that sense anymore. This poses difficulties to application of OLS. However, under certain conditions it is still possible to apply OLS to time series, then it is called "time series regression."
So, how do you apply cross-sectional tools to time series? You convert the time series data into the format that is expected by cross sectional tools. Suppose, that you think the sales are according to the following model:
$$S_t=c+\phi S_{t-1}+X_t\beta_t+\varepsilon_t$$
In this case you re-write it like this:
$$S_i=X'_i\beta'_i+\varepsilon_i$$
Here you have design matrix matrix $X'_i=[1, S_{i-1}, X_i]$, that contains with the predictors a column of past sales of the store. This allows us to apply the OLS to time series.
One thing to note that as I wrote earlier simply applying cross-sectional tools to time series problem doesn't make it right to do so. When you recast your problem as cross-sectional you must ensure that the assumptions underlying the tool are still valid, and this is not always the case.
